Skip to content

Changelog

0.24.1

A docs release: the About page shows NuBoard's main features in tabs, a new guide covers sharing results from a cloud machine, and the Python API pages have more pictures. No code changes (#37).

  • Feature tabs on the About page: the single screenshot is now six tabs (Overview, Ranges, Timing, Audio, Timeline and Hypotheses), each with a screenshot in light and dark and a short description in a box of fixed height, so switching tabs never moves the image. Nothing rotates on its own.
  • Sharing from a cloud machine: a new guide compares SSH port forwarding, a token on the network and HTTPS behind a reverse proxy, with the pros and cons of each for accessibility and security. The viewer guide and the command line reference link to it.
  • More pictures in the Python API pages: add_scalar logs eval/loss with low and high and shows its band, add_scalar_baseline and timer show what their examples produce, and the add_audio example logs a plucked arpeggio with harmonics and vibrato instead of a chirp, for a richer spectrogram.

0.24.0

Artifact versions can be made from a run's files in the viewer, and an artifact's files are read by their path inside the version. Artifacts can no longer contain other versions, and the Artifacts page no longer hangs on a bad version. No breaking changes; the viewer now asks before adding files identical to the latest version.

  • Reading an artifact's files: an Artifact reads like a mapping from each listed file's path inside the version to the file, so art["best.pt"], "best.pt" in art, list(art) and len(art) work for copies and references alike. Only the files the manifest lists are keys, which matters for a reference to one file, whose files folder may hold others.
  • Artifact versions can no longer contain other versions: a source inside a project's artifacts/ folder, or one that contains it or another project's versions, is refused. Before, referencing the project folder made each new version list the previous versions' manifests.
  • The viewer asks before adding identical files again: adding or promoting files that are the same as the artifact's latest version shows "the same as v1" with Add anyway (allow_duplicate in the API) instead of adding a duplicate version.
  • The Artifacts page cannot hang on loading: a version with a missing, unreadable or malformed manifest is listed with the problem instead of breaking the list, and a list that fails or takes over 20 seconds shows an error with Retry. A copy that is interrupted removes its staging folder, and staging folders abandoned for 6 hours are removed when a version of that artifact is added or the viewer starts with --edit-artifacts.
  • Artifact versions from a run's files: with nuboard up --edit-artifacts, the Files tab has New artifact version next to Download, and an artifact icon on folder rows. It opens the same form as New version on the Artifacts page, with the source filled in and a choice of Copy (the default) or Reference. A reference to files inside the project is stored relative to the project (reference_base: "project"), so it survives copying, rsyncing and merging; a running run can only be copied. The run is recorded as the producer and the run folder is never changed. Without the flag the buttons are hidden. nuboard artifact add with a path inside a run does the same, promote_run_path does it from Python, and nuboard rm keeps a run that a referenced version needs unless --force.
  • Run details tabs have icons, and only the first five show in the row: the rest, and the Charts, Sweeps and Media shortcuts, are in a ⋮ menu, so the settings filter stays on the same line on a laptop screen.
  • Timing tab: the dot marking the shown step on each per-step line is round (it was stretched with the line), and a timer's full chart marks the shown step and follows the cursor with a dot.

0.23.0

Timers measure where training time goes, with a new Timing tab in the run details, and the Charts alerts follow the legend. No breaking changes.

  • Timers (#32): run.timer(name) times parts of the training code (an epoch, validation, a callback) as a context manager or with start() and stop(), which can be in different methods: the same name returns the same timer. accumulate=True sums spans (such as every batch) into one point per flush(); sync=True waits for CUDA; only rank 0 logs unless all_ranks=True. Misuse warns once and never stops training, and a timer logging many points per second suggests accumulate. nb.Timer.disabled() measures without logging. Timers are scalars under timing/; init(timing_prefix=..., system_prefix=...) renames the run's own prefixes (recorded in run.json), and a scalar logged by hand under one of them gets a one-time warning.
  • Timing tab in the run details: calls, mean, p95, total and share per timer with a per-step line and a full chart, nested timers, time spent waiting for other ranks and not under any timer, findings (spikes, a slow rank), and a timeline of one step with a lane per rank. On the Charts page the timing and system sections come last, folded, and timers are drawn as durations.
  • Charts alerts follow the legend (#33): the alerts strip under the focused chart lists only the runs switched on in the legend (with grouping, the runs of the visible groups). Alerts of runs switched off are counted in a "+N in hidden runs" note, which switches them back on when clicked.

0.22.0

A new Hypotheses page keeps the story behind your runs, and times can be shown in a time zone you pick. No breaking changes.

Times can be shown in a time zone you pick, and each run records the zone of the machine(s) it ran on.

  • Time zone chip in the header (the clock button), optional: it reads Local time by default, showing times as before in your browser's zone. Pick Machine time to show each run on the clock of the machine that recorded it (axes shared by runs from different zones switch to UTC and say so), or UTC, a named zone such as America/Santiago (daylight saving followed at each moment) or a fixed offset such as UTC-03:00. It changes only how times are written: data, queries, links and saved views are unchanged. All views use it at once (run table, run details, chart tooltips and wall-time axes, the Timeline, artifact and file dates, exported reports), and hover read-outs end with the zone's short name. The choice is kept per browser. See Time zones.
  • Each session records its machine's zone: sessions entries in run.json gain utc_offset, utc_offset_end and host, so a run resumed on a machine in another zone shows each session's offset in Run details (machine_zone, plus machine_zone.s1, … for resumed sessions). Not a breaking change: the keys are additive, and older runs show the first session's offset worked out from the run id, or unknown.

A new Hypotheses page keeps the story behind your runs.

  • Hypotheses page (#28): keep track of what each run was testing and which run led to the next idea. A project can have several boards; each draws its hypotheses as a tree of cards (with verdicts, run outcomes, notes, linked reports and the reason on each arrow) or as a timeline. Runs and reports that are gone show as "not found" instead of breaking the board. Boards are plain JSON in <project>/hypotheses/<board>/, safe to commit; see Hypotheses and the file format. The demo's vit-ablations project has one. --read-only-reports now also makes boards read-only.

  • Runs selection bar: Compare (or Details for one run) moved into the blue bar with the other actions on the ticked runs; Plot selected is gone (the bar's Charts does the same); Tags…, Set property… and Add to collection… are now Tags, Properties and Collection.

0.21.0

init no longer hangs a multi-GPU job when only rank 0 calls it, and a new Benchmarks page compares nuboard with TensorBoard. One breaking change: init(gpus=...) is now system_devices.

  • Breaking: init(gpus=...) is renamed system_devices, next to system_metrics and system_interval, so it reads as the devices system metrics track, not the devices you train on. Replace gpus= with system_devices=; the old name is gone.
  • init no longer hangs a multi-GPU job when only rank 0 calls it, as is usual with TensorBoard (if rank == 0: run = nb.init(...)). Rank 0 used to broadcast its run id through torch.distributed and wait for every rank to receive it. It now never waits for other ranks: they find the run through a shared job id (SLURM_JOB_ID, or torchrun's --rdzv-id), and without one they log nothing, which loses only other machines' system metrics (#24).
  • Benchmarks page in the docs: nuboard against tensorboardX and torch.utils.tensorboard, as interactive charts (exact values and ratios on hover), with the script that reproduces it (benchmarks/vs_tensorboard.py).
  • Full screen replaces Expand in Files. A text file taller than its 520 px box opens full screen, as charts do: it fills the screen and scrolls inside however long it is, and Exit full screen or Esc brings it back. The page around it no longer grows with the file, so a huge log can't stretch it. A file that fits its box has no button.
  • Only changes and the Diff … against … run menus are also in a full screen diff, so a long file compared between runs can be cut down to the lines that differ without leaving it.

0.20.1

The Files tab no longer jumps back to the top while a run is live.

  • The Files tab keeps its scroll on live runs. A running run's heartbeat no longer reloads the open file, so a long file such as a config.yaml stays where you scrolled instead of jumping to the top. A text file the run is still writing is read again as it reports in and grows in place, following new lines only when you are at its end (#21).
  • Expand next to a text file in Files or a pinned tab shows it at full height instead of in a 520 px box; the choice is remembered (#21).
  • The demo's live run saves a config.yaml and writes a growing files/progress.log.

0.20.0

The first release on PyPI: pip install nuboard (or uv pip install nuboard).

  • nuboard rm deletes runs chosen by a query (-q 'status = crashed') or by id or name (--run, repeatable). It lists the runs with their sizes and deletes nothing until you confirm; --dry-run and --yes work as in merge. Running runs are skipped unless --include-running, artifact versions the runs produced are kept, and with neither a query nor --run it refuses, so it never deletes every run by accident (#18).
  • Command line reference. A new page lists every command with its options and defaults (#18).
  • Known limitations have their own page.
  • Viewer pages can't be framed by another website, so a page you visit cannot overlay the viewer to trick clicks, and the sign-in form no longer follows a next link with a backslash to another site.
  • pip install nuboard includes the viewer. NumPy, PyArrow, Starlette and uvicorn are now regular dependencies, so nuboard up works without [viewer]; nuboard[viewer] still installs. Logging still uses only the standard library and imports none of them.
  • Package metadata for PyPI: project links and classifiers. The docs are published at https://eagomez2.github.io/nuboard/, and GitHub workflows run the tests and publish the docs.

0.19.2

  • nuboard merge SRC DEST merges a project logged on two machines into one folder. It shows the full plan first (runs to copy, identical items, conflicts, renamed or renumbered items) and writes nothing until you confirm. It copies what DEST lacks and never overwrites or deletes: clashing artifact versions are renumbered and copied runs repointed, collections and hide rules are combined, and a report or view that differs is copied next to DEST's. --dry-run, --yes, --json, -v, --project and --include-running (#16). See Working on several machines.

  • TensorBoard import: videos and meshes. import-tensorboard now imports add_video clips (animated GIFs) as videos and add_mesh summaries as point clouds with their colors, instead of an image and a skipped plugin.

  • Small charts show values on hover. Hovering a chart in the All scalars grid shows each run's value at the cursor, best first, up to eight and "+N more" (#3).
  • A pinned step shows on every chart. While the main chart is pinned, every small chart draws a dashed line at that step with a dot on each curve, and hovering one shows its values there (#3).
  • Full scalar names under small charts, wrapping instead of cut off; Full names in the display row turns them off (#3).
  • Timeline hover cards in reports. A Timeline snapshot in a report shows the Timeline page's hover card on each bar (name, id, status, project, start, duration, alerts) in the preview and in exported HTML, instead of a slow browser tooltip; in NuBoard, clicking a bar opens the run (#3).

  • More room for reports. A report with text opens in Preview (the view you pick is remembered). The report list and the Workspace list on the Runs page fold into a thin bar with their button or [, remembered per browser (#3).

  • Datasets as Sweeps parameters. The artifact versions each run used are parameters named inputs.<artifact> (v1, v2), so runs can be compared by data version without copying it into the config (#14).

  • Readable list and dict values in Sweeps. Axis labels, scatter labels and the Best runs table show only what differs between values ([0]=0.08, flip, mixup, [1].dim=512), with the full value on hover and, on click, pretty-printed with a Copy button (#14).
  • Split a list into element axes. A list of plain values with one length in every run can be split in the picker into key[0], key[1], ..., kept in the URL as split= (#14).
  • The demo shows them. Half of the llm-pretrain sweep trains on the older fineweb-edu:v1 (a little worse), and vit-ablations runs have list-valued transforms and crop_scale settings. Delete an existing demo folder to get the new data.

0.19.1

  • Sweeps parameter picker. Parameters opens a searchable list of every differing key, grouped by prefix, with its type, number of values and correlation with the objective. Chosen keys are chips you can drag to reorder and remove with ×. Clear all and Reset to defaults are in the list.
  • Better default parameters. With nothing picked, Sweeps shows up to six keys, the most correlated with the objective first, leaving out seeds, paths, id-like keys, long text and keys with many values. The last choice is remembered per project.
  • Hide rules for Sweeps. A project can hide keys from the picker with a filter such as !seed !.path, saved in <project>/sweeps.json.

  • Categories with many values keep their 11 most common values on the axis and gather the rest in (other).

  • nuboard demo no longer piles up runs. Run again on a folder that already holds the demo, it opens it instead of adding 31 more runs each time, so it starts in about the same time on every run. Delete the folder for a fresh demo.

  • Open selected runs in Charts and Sweeps. The Runs selection bar has Charts and Sweeps next to Media. Charts, Sweeps, Media and Timeline show a Showing N selected runs banner while limited to a selection, with buttons to the other pages (which keep it) and Show all; Clear selection also drops it from the URL (#7). A run opened for inspection (or the runs being compared) has the same Charts, Sweeps and Media buttons after its tabs.

  • Report picker. The Report field of Add to a report (every bookmark button, and Add N to report on Media) lists the project's reports as soon as it is focused, newest first, with their titles and file names. Typing narrows it by title or name with the matches highlighted, ↑/↓ and Enter pick, the last used report is highlighted (and taken when nothing is typed), and a new name is offered as New report "name". It lists the newest 50 and counts the rest, which typing reaches (#12).

  • Audio in reports. Insert → Audio (or dropping an audio file on the text box) uploads a WAV, FLAC, MP3, Ogg, M4A or WebM file, checked from its bytes and capped at 50 MB (--max-audio-mb), and writes [caption](files/<hash>.wav). It plays with the waveform player in the preview, is copied by the .zip export and embedded in the HTML export with the player chosen there (#12).
  • Report toolbar icons. Link, lists, checklist, quote, code block, rule, table, undo and redo are drawn as theme icons instead of emoji and symbols, so they follow the light and dark themes and the pressed color (#12).

  • Media item is grayed out in Insert when no run of the report's project logged media, saying so. Runs now record media_count in run.json with each write they already make; runs from older versions are checked once for a media pack when indexed (#12).

  • Reproducibility. A new guide explains what a run records, how to replay it, and what is not covered. init now:

  • removes any user, password or token from the recorded git remote URL;
  • records the repository of your code: code_dir= if given, else the running script's folder, else the working directory (git.source), with a warning when the script and the working directory are in different repositories;
  • records git.pushed (whether a remote-tracking branch has the commit) and git.untracked (how many untracked files, never their names), shown and highlighted in All settings;
  • keeps the commit of packages installed from git in requirements.txt, without credentials;
  • records the folder and git state of packages installed from a local folder (pip install -e ../lib): a clean, pushed checkout becomes an installable git+URL@commit line, uncommitted changes are saved as files/git-<package>.patch (#13);
  • never lets a repository's config run a program during capture (core.fsmonitor off, diffs with --no-textconv), and skips patches over 10 MB, recording git.patch_bytes instead (#13);
  • records what it can in a repository with no commits, on a detached HEAD, from a deleted working directory or with broken package metadata, and never fails the run over it (#11).
  • Report editor: formatting toolbar and Insert menu. Buttons for bold, italic, underline (<u>), strikethrough, headings, lists, checklists, quotes, code, links, rules and tables write plain Markdown, with Ctrl/⌘+B, I, U, Shift+X and K, undo and redo, and show as pressed when the cursor is inside their formatting (pressing again removes it). Insert adds an Image (pick, drop or paste, with alt text and a caption), a Text file (inline as Markdown, a table or code, or attached as a collapsible block with a download link), a Table, or goes to Charts and Media. Dropping or pasting an image into the text box uploads it too (#4).
  • Uploads are checked and stored in the project. They go in reports/files/, named by content hash, and are part of the .zip and HTML exports. The server checks the bytes (PNG, JPEG, GIF, WebP or a plain SVG; UTF-8 text with no NUL bytes), caps sizes at 20 MB for images and 5 MB for text (--max-image-mb, --max-text-mb), and serves report files with a fixed type, nosniff and a strict CSP; text files always download. nuboard up --read-only-reports turns off every change to reports.
  • Report images from other websites are not loaded; they show as a link.
  • Filter the settings of compared runs. A filter box in All settings narrows its rows by key: plain text, /regex/ (or the .* toggle), !exclusions and several terms at once, optionally matching values too. It works with Only changes, shows "12 of 340 keys", highlights the matches and is kept in the URL. / focuses it and Esc clears it (#9).
  • Sweeps sees every config key. A key that differs between runs gets a parameter chip however deep it sits in a large config. Sweeps read its parameters from the run list, which keeps only the first 64 config values of each run, so later keys (such as generator.config.weighted_sir_db.edge.min) were missed. It now asks the new GET /api/sweep-keys?runs=... endpoint, which returns every differing key with each run's value; the run list stays small.
  • The demo shows it. speech-denoise runs now carry a long training-mix generator config, and differ in generator.config.weighted_sir_db.edge.min, its 73rd value: open Sweeps on that project to see the chip.

0.19.0

  • Download a waveform figure's audio and signals. Waveform figures (nb.WaveformFigure) have a Download menu next to Fit, on the Media page, in the large view, in report embeds and in exported HTML: Audio (.wav) saves the audio exactly as logged, and Data (.zip) a bundle of audio.wav, one float32 .npy per signal and a metadata.json with the sample rate, signal names, range, regions, caption, notes, run, key and step. The raw item, which no other tool could open, is no longer offered.
  • The report .zip export writes the same bundle for a waveform figure, instead of a .wav and a -signals.csv that lost the regions and range and grew to millions of rows for long clips.
  • nb.WaveformFigure.load(path) rebuilds a figure from a bundle; logging it again stores an identical figure. run.waveform(key, step) in the analysis API returns (audio, sample_rate, signals, meta) without downloading.
  • Other media downloaded from their card get a file name extension from their type (.png, .wav, .json, ...).

0.18.0

  • Interactive charts in reports and exported HTML. A chart saved from the Charts page lists every series' value at the step under the pointer, best first, as the Inspector does, in the viewer and in the exported HTML file; a click pins the values. In the exported file, PR curves show the nearest threshold with its precision, recall and TP/FP/TN/FN counts on hover, and distributions the bin under the pointer for each run, as in the viewer.
  • The all-runs chart in reports. The "All runs on one chart" overlay on the Media page can be added to a report, where it is drawn live from the logged items (hover, legend, Open in Media), and exported to HTML with its hover read-outs. It also downloads as a PNG with its legend.
  • Timeline snapshots are SVG. A Timeline added to a report is saved as an SVG drawing, sharp at any size, with a hover title on every bar (name, id, status, start, duration, alerts) in the viewer and in exported HTML.
  • Best-step mode on the Media page always shows the choice last made: an answer for an earlier choice (Lowest before Highest, another scalar, a live refresh) could arrive later and show the other end. The line under the cards now says Lowest or Highest, and notes when the nearest logged item is from a different step than the best value.
  • Open the all-runs chart large. The "All runs on one chart" overlay of PR curves and distributions on the Media page has a magnifier button that opens it in the large view, as single cards do; Esc closes it. It can also be downloaded as a PNG, with its legend, or added to a report as a snapshot with an Open in Media button.
  • Run ids on the Timeline. Each row shows the end of the run id after the name, so runs with the same name can be told apart, and hovering a bar shows the full id.
  • The AP next to a PR curve says what it is on hover (average precision, the area under the curve), and the docs say how it is computed.
  • Space plays and pauses audio in full screen. Waveform figures and the audio player toggle playback with Space while in full screen, and only there: outside full screen, while typing (such as a note) or with Ctrl, Alt, Cmd or Shift, Space does what it did before, and Esc still leaves full screen. The audio player used to take Space whenever it had focus; it no longer does outside full screen.
  • Artifacts with many files stay fast in the viewer. The Artifacts page no longer loads every version's file list: each version comes with its file count and size, and its files load 1,000 at a time when you open it, with Show more and a path filter. Before, one version with a million files slowed the whole page. The New version form warns from 10,000 files and suggests referencing shards or an index file instead.
  • New: nuboard artifact unlink and nuboard artifact restore. They unlink a version (or with --all, every version) and restore one, as the viewer does: nothing is deleted. nuboard artifact ls --unlinked lists unlinked versions too.
  • Docs: Recommendations for registering data: shards over millions of files, when to hash, versions as snapshots. The guide no longer says references are checked by modification time: they are checked by size.
  • New: init(gpus=...) picks the GPUs that system metrics track, by the device names your script uses: gpus="cuda:0", a list such as ["cuda:0", "cuda:1"], a torch.device, "mps" or "all". A name that cannot be tracked raises ValueError. See Choosing GPUs.
  • Changed: GPU metrics respect CUDA_VISIBLE_DEVICES. Only the GPUs the process can use are tracked, named by the index the script sees, so with CUDA_VISIBLE_DEVICES=2 the machine's third GPU logs as system/gpu0/.... Without the variable nothing changes. run.json's env.gpus lists the same GPUs, now with device, index and uuid.
  • Breaking: all_ranks is removed from init, add_scalar and add_text. In distributed training only rank 0 logs, and the calls are no-ops on other ranks. The viewer only plots rank 0, so values from other ranks could not be seen anyway. System metrics are still sampled by local rank 0 of each machine. Drop the argument from your calls; runs logged before keep their per-rank files and still load.

0.17.1

  • Fixed: a pinned point now follows the chart when you zoom or pan. In Charts, the dots of a pinned point kept their place on screen while the axes changed under them, and disappeared once the data for the new range loaded. They now stay on the points they mark.

0.17.0

  • Breaking: init(name=...) is now init(run_name=...), so it reads as the pair it forms with run_id. run_name is a free label shown in the viewer; it is no longer sanitized, since it is not part of any path.
  • New run ids. The default run_id is <YYYYMMDDTHHMM>-<hash>, such as 20261001T1833-48ad8bb2: a local timestamp plus 8 random hex digits, redrawn if the folder already exists. The run name is no longer part of the id. You can still pass your own run_id; existing runs keep their ids. TensorBoard imports use the same shape, with a hash of the source folder.
  • Charts: a Ranges switch hides the shaded bands (group mean ± std or min–max, logged low/high bounds, rank ranges) when they clutter the chart. It is saved with charts added to reports. Smooth and the display switches moved to their own row under the grouping and x-axis controls.
  • Charts: Ctrl/⌘ + mouse wheel (or a trackpad pinch) zooms the focus chart around the cursor, also past the ends of the data, with y fitted to the view; Alt + wheel zooms both axes; Shift + drag pans in any direction and Shift + wheel along x. The view stays put while finer data for it loads, and the data comes with a margin on each side so smoothing does not restart at the edge. Dragging still zooms to a range and double-click still resets.
  • Charts: the inspector's latest values leave out hidden series, and take the last value in the zoomed range.
  • Timeline: Day / Week / Month / Year / All windows with ‹ › to step through time, Ctrl/⌘ + wheel zoom around the cursor, Shift + wheel or Shift + drag to move, and month and year ticks on long ranges.
  • Calmer live updates. Charts redraw new points only while you are not zoomed, pinned or hovering, with a Live switch and a New data · Update button otherwise. That button and Reset zoom sit in the chart's fixed-height header, a slow fetch shows a small spinner instead of swapping text, and metric tabs stay on one row, so nothing moves when data arrives. Runs table rows hold still under the pointer. Media, Sweeps, Timeline, Artifacts and Reports no longer change under you: a New results · Refresh button appears instead, and heartbeat-only changes are ignored.
  • Audio players label the waveform's amplitude (±1 full scale, per channel), with an Amplitude axis switch on the Media page.
  • Waveform figures (signals and region lanes, such as VAD) get a full screen button like the audio player's, in Media and in exported reports.
  • Run details follow the run's status, duration and summary live, so a run that finishes shows Finished there too.
  • Tables: a highlighted row no longer gets a rounded corner when it is the last row; table frames have square corners.
  • Updated the screenshots in the docs.
  • Waveform figures. run.add_figure(key, nb.WaveformFigure(audio, sample_rate, signals)) logs audio with signals over its waveform, one value per sample, such as a voice activity probability or a gain mask. The viewer draws it as a plot that is accurate at every zoom (M4 decimation: the first, smallest, largest and last sample of each pixel column, recomputed as you zoom), fast on files an hour long, down to single samples. Hovering shows the exact values at a sample; the clip plays from where you click. It is listed under Figures; add_audio and its spectrogram player are unchanged. regions= adds labelled tracks from training (a list of segments, a class per frame, or multi-hot frames for classes that overlap), drawn in lanes under the plot. Dragging selects a section to loop, zoom to or note; notes are kept apart from the logged regions, in the run's annotations.json. In a report, an HTML export looks and works the same, offline, with its audio, regions and notes, and a .zip export writes its audio as a WAV and its signals as a CSV. See add_figure.
  • The demo's speech-denoise runs log vad/enhanced, a waveform figure of the enhanced output with a speech probability and target, predicted and overlapping event tracks, with notes on the first one's false alarms; its listening report embeds one.

0.16.0

  • Jobs. A second process can log into a running run with nb.init(..., run_id=ID, job="eval"), for example an evaluator that scores checkpoints while training goes on. Its scalars share the run's step axis, and it never changes the run's status, heartbeat or summary: its own state goes in jobs/<job>.json, its console in files/output.<job>.log and its system metrics under system/<job>/. The viewer shows each job as a chip next to the run. A job can only join an existing run, the same job cannot run twice at once, and logging a name the main job also logs gives a warning. The last job to end compacts the run. See Evaluating while training, which includes a checkpoint queue recipe. Runs without jobs are unchanged, and older versions still read new runs (without the jobs).
  • Storage. A disk button in the left rail shows how much space runs, artifacts and other files take, per project and per run. The Artifacts page has the same button for artifact versions. It is measured only when you press Measure or Refresh, pauses after 10 s on large folders to ask before going on, and counts only what is stored in the folder: a referenced artifact counts its manifest, with the size it points to shown apart.
  • Sweeps: the scatter plots under the parallel-coordinates chart no longer scale their axis labels with the plot width; text keeps the same size with one column or four.
  • The demo's convtasnet run has an eval job that logged eval/dnsmos.

0.15.8

  • Access token off localhost. nuboard up --host with an address other than 127.0.0.1 now protects the whole viewer with a token, like Jupyter: it prints a link with a random token, and opening it signs the browser in. Without the token, pages show a sign-in form and API calls get a 401. --token VALUE or NUBOARD_TOKEN sets a fixed token and --no-token turns it off. Scripts send Authorization: token VALUE. On 127.0.0.1 nothing changes. If you already serve on 0.0.0.0, open the printed link once, or pass --no-token to keep the old behavior.
  • Changes sent from a page of another website through your browser are refused, with or without a token.
  • Adding artifact versions in the viewer. Started with --edit-artifacts, the Artifacts page has New version: a form that adds a version by reference to a file or folder (absolute, ~, $VAR or relative to the served folder), checks the path as you type, and shows a review step before adding. Nothing is copied or uploaded.
  • Unlinking. With --edit-artifacts, Unlink version moves a version to artifacts/.unlinked/ after a confirmation, without deleting any file. Show unlinked lists them, and Restore puts one back. Runs show such a version as "(unlinked)", and version numbers are never reused.
  • Without --edit-artifacts, artifacts are read-only in the viewer, as before.

0.15.7

  • New run.add_scalar_baseline(key, value, *, low=None, high=None) sets a reference value for a scalar, such as the score of the unprocessed input: run.add_scalar_baseline("val/pesq", 1.31). It is linked to the scalar by name, saved in run.json right away, and replaced by calling it again; a dict sets several at once.
  • Charts draw it as a dashed horizontal line, with a light band for low and high. Runs with the same baseline share one gray line; a baseline only one run has is dashed in its color. Hiding or highlighting a run does the same to its baseline, grouped charts use each group's mean baseline, and the y-axis keeps baselines in view. The inspector lists each baseline and its runs.
  • A Baseline switch next to Points turns them off. It shows only when a run has a baseline, so other charts look as before. Charts saved to reports and exported to HTML keep the setting.
  • The demo's speech-denoise runs have PESQ and SI-SDR baselines.

0.15.6

  • add_scalar takes optional low and high, a range around the value such as mean ± std over an epoch: run.add_scalar("val/loss", mean, step=epoch, low=mean - std, high=mean + std). Charts shade the range around each run's line, and the inspector shows it after the value. Scalars logged without them work exactly as before. When runs are grouped (for example several seeds), the band shows the spread across runs as before.
  • The demo's llm-pretrain runs log eval/loss with a range.

0.15.5

  • Console logs show their colors. Color codes (the 16 terminal colors, 256-color and true color, bold, dim, italic, underline, inverse) are drawn instead of removed, with terminal colors that suit the light and dark themes. Raw shows the text as stored. The log file is unchanged.
  • In the Files tab and pinned tabs, any text file with color codes or progress bars (such as output.log or your own .log files) is drawn the same way instead of showing the codes, and two runs' logs are diffed by their visible text, so a change of color alone is not a difference.

0.15.4

  • Labels start with a capital letter everywhere in the viewer: toggles, buttons and menus (Best / Last, Show archived, No grouping, Wall time, Ignore outliers, Log y, Pick step, Waveform, One row per run, Packed, ...), hints, tooltips and run statuses (Finished, Running, Crashed). Only the first word is capitalized. Names that come from your data (scalars, config keys, run names, tags, captions) are shown as written, and queries still use lowercase statuses (status = "finished").

0.15.3

  • Files tab: the tree is on top, full width, about 12 rows tall with its own scroll bar and a corner to drag it taller; the preview and diffs use the full width below it. Long names lose their middle rather than their end, and the full path shows on hover. files/ starts open, other folders folded.
  • The row of tabs no longer has a tab for every file in files/; Console stays. Instead, any file can be pinned as a tab, by path or by a pattern such as checkpoints/*/best_*.pth, which picks each run's own match (newest first) so differently named files are compared side by side. Pins are shared by everyone through <project>/pins.json.

0.15.2

  • Run details and Compare have a Files tab that lists every file in the run folder as a tree, with the number of files and their size per folder. NuBoard's own data files are hidden unless asked for. Text files (logs, YAML, JSON, CSV, patches, code) are shown, up to their first 2 MB, and diffed side by side between two runs; images are shown; every file can be downloaded.
  • Checkpoints (.pt, .pth, .ckpt, .bin, .safetensors, .npy, .npz) list their keys as a tree with each tensor's shape, dtype and size, plus the total parameter count. Between runs the keys are diffed (added, removed, changed shape or dtype). PyTorch is not needed, and no code in a checkpoint ever runs: the pickle is read with every class replaced by an inert placeholder. A malformed, cut-off or hostile file shows what could be read and why the rest could not, and never stops the viewer.
  • Compare: the All settings table has the same borders as the other tables. Changed rows were picking up the file diff's styles.

0.15.1

  • Ctrl+C stops nuboard up at once, even with a viewer tab open. Before, the open live-update connection kept the server waiting forever.
  • The run table never cuts values off: every column is as wide as its widest value, and a table wider than the window scrolls sideways. All config keys that differ between runs are shown, not only the first eight. Notes are the one column still capped; their full text is in the notes editor.
  • The Columns menu has a search box, so a column is easy to find among many scalars and config keys.
  • Charts: the big chart has an All N picker that searches every scalar, so any of them can be brought into focus. With more than 12 series, the legend shows the first 10 and a "+N more" button, plus a filter that matches run names and config values (lr=3e-4, layers=24); Only these hides every series that does not match.
  • Charts: scalar sections fold like Media sections, with Expand all and Collapse all. With more than six sections only the first three start open; what you open or close is remembered.

0.15.0

  • nuboard artifact add adds an artifact version from the command line. Give every value as a flag, or run it in a terminal and it asks for what is missing, one field at a time. Each answer is checked as you enter it (the project and name, that the path exists and holds files, the store, the type, a non-empty description), < goes back a field, metadata can be added, edited and deleted with its type shown, and nothing is written until you confirm the review. Without a terminal or with --no-input it never asks; every invalid or missing flag is reported at once. --dry-run checks without writing.
  • nuboard artifact ls lists artifact versions, and nuboard artifact show prints one and checks its files.
  • Artifact versions have a description (add_artifact(..., description=)), and the manifest records who created each one (created_by: user, host and command). The Artifacts page shows both.

0.14.1

  • A leading ~ is expanded in artifact paths and in NUBOARD_DATA_DIR, so nb.add_artifact("datasets", "fineweb-edu", "~/data/fineweb-edu", reference=True) works as written.

0.14.0

  • Reference artifacts: add_artifact(..., reference=True) records where a file or folder is, with each file's size and modification time, and copies nothing, for datasets that live on shared storage. checksum=True also hashes every file. use_artifact warns when a version's files are missing or changed in size since it was made, and Artifact.verify() lists them.
  • References work across machines: a path inside $NUBOARD_DATA_DIR, or one written with any environment variable ("${IMAGENET}/train"), is stored with the variable and expanded where the artifact is used, so one version points at the laptop's copy on the laptop and the cluster's on the cluster.
  • Artifacts across projects: run.use_artifact("datasets/fineweb-edu:v2") uses an artifact of the datasets project from a run in any other project, and run.add_artifact("datasets/fineweb-edu", ...) adds one there. The run's inputs and outputs record the other project (project, project_path). An artifact name can no longer contain /.
  • nb.add_artifact(project, name, path, ...) adds a version without a run, such as a dataset no script produced, and nb.get_artifact(ref) finds one without recording it, such as in a notebook.
  • Artifacts page: projects that hold only artifacts are listed, a project's list includes the artifacts its runs used from other projects, references show where their files are, and the use_artifact line names the project. Run details show another project's artifact as project/name:vN.
  • The demo keeps its datasets in shared-data/, registered as references in a datasets project that llm-pretrain and vit-ablations use.
  • The Artifacts guide now explains how to organize datasets, checkpoints and projects.

0.13.3

  • Fixes NotADirectoryError: ... runs/.DS_Store/run.json on macOS: a file where a run, project or artifact folder is expected, such as the .DS_Store that Finder leaves in folders it has shown, is now skipped. This broke nuboard demo on a folder that already had demo runs, and listing artifacts with a stray file in artifacts/.

0.13.2

  • Fixes pnpm install in frontend/, which failed with "No matching version found for eslint@^10.13.1": the 0.13.0 version bump had also changed eslint's version range in package.json. It is back to ^10.12.0, as in the lockfile, and a test now checks that package.json and pnpm-lock.yaml agree and that every version string matches.

0.13.1

  • Saved views can be deleted from the Runs page: hover a view in the workspace list and click ×, then confirm. The view's file in <project>/views/ is removed; runs, collections and reports are not touched. The × of views and collections also shows when the row has keyboard focus.

0.13.0

  • Breaking: run.log is removed; run.add_scalar is the one way to log scalars. Replace run.log(data, step=s) with run.add_scalar(data, step=s), which takes the same dict and keyword arguments. The docs, examples, tests and benchmarks use add_scalar throughout.

0.12.0

Logging

  • run.add_scalar logs scalars, named as in TensorBoard: run.add_scalar("train/loss", 0.42, step=10) logs one, and run.add_scalar({"train/loss": 0.42, "train/lr": 3e-4}, step=10) several, which covers TensorBoard's add_scalars too. run.log keeps working as the same call (removed in 0.13.0).
  • run.add_video logs clips from (T, H, W, C), (T, C, H, W) or (N, T, C, H, W) arrays and tensors, encoded as looping animated PNGs with no ffmpeg needed, or MP4, WebM, MOV, Ogg and GIF files as they are. The Media page has a Video kind with a frame player, and the ppo-atari demo logs rollout videos.
  • System scalars: nb.init samples CPU, memory and GPU use every 30 seconds into system/ scalars, from LOCAL_RANK 0 of each machine. GPUs are read with nvidia-smi, and PyTorch's CUDA and MPS allocators are read when the script has imported torch. Turn it off with system_metrics=False, pick groups with system_metrics=["cpu", "memory", "gpu"], and change the rate with system_interval.
  • Console capture: what the script prints to stdout and stderr, including the traceback of a crash, is also written to files/output.log and shown as Console in the run details. Turn it off with capture_console=False.
  • PyTorch tensors passed to add_image, add_audio, add_histogram, add_point_cloud and add_video are copied to NumPy before the call returns, so changing the tensor in place afterwards no longer changes what is logged. bfloat16 tensors work.

Naming

  • Docs, labels and the Python API say "scalar" instead of "metric", as TensorBoard does: All scalars, X axis scalar, best scalar. RunReader.scalar_names replaces metric_names, which stays as an alias.
  • Breaking: RunHandle.scalars() and Project.scalars() in the analysis API name their first column key instead of metric.
  • The file format is unchanged, and so are saved views, report charts and query syntax.

Docs

  • The Python API section has a page per method, all listed in the menu, including add_video and the new nb.init options.

0.11.0

Runs table

  • Tags can be edited in the viewer. In a run's details, the tag button under its name removes a tag with × and adds one by typing, with the project's tags suggested. Tags… in the selection bar adds and removes tags on every selected run, keeping each run's other tags. Once edited, a run's tags replace the ones it logged, so those can be removed too; run.json keeps the logged ones.
  • Custom fields and properties are now one thing, properties. Set field… is now Set property…: it picks a property, then its value with a control that fits its type (buttons for a choice or yes/no, pills for tags, a date picker), and a new name is defined with its type first. Columns lists every property of the listed projects under Properties, marks values without a definition as untyped, and New property… defines one and shows it as a column. In a run's Properties tab, untyped · define gives such a value a type. The stored format is unchanged.

Docs

  • A Python API section describes every call, in the style of PyTorch's SummaryWriter reference: its signature, parameters, the shapes and dtypes it accepts, what it stores and how big, what the viewer shows, a complete example and a picture of the result. It ends with a table mapping each SummaryWriter method to NuBoard.
  • Get started walks through training a small NumPy network on a spiral with three learning rates, logging scalars, images, PR curves, distributions, a point cloud and a table, then reading them in the viewer. The script is examples/get_started.py.
  • The examples in the API pages run as tests, so they stay correct.

0.10.2

  • The Columns menu of the runs table floats above the page, so a table with only a few runs no longer cuts it off.

0.10.1

  • HTML export draws audio with the viewer's waveform and spectrogram player, embedded once in the file and working offline, with notes on its lane and the page's light or dark theme. When a report has audio, the export dialog offers Audio player: Waveform (default) or Simple, the browser's own player as before.

0.10.0

Media

  • Distributions: run.add_histogram(key, values) stores bin counts and the values' min, max, mean and standard deviation, as TensorBoard's histograms do. They show with the figures: each card draws the chosen step filled with the step before it dashed, hovering shows a bin's share, and over steps shows the median and 68%, 86% and 100% bands across training. A chart above the cards overlays every run. They export to HTML, and the vit-ablations demo logs weight, gradient and logit-margin distributions.
  • nuboard import-tensorboard imports histograms (TF1, TF2, PyTorch and tensorboardX), re-binned into 50 even bins.
  • Point clouds: an axes button draws x, y and z axes with round values. add_point_cloud(..., axes=True) shows them from the start, and three names label them, as the demo's embeddings do.

0.9.1

  • The run details panel (click a run name) shows the run's tags under its name, and updates when they are edited.
  • Roadmap: dropped the finished TensorBoard import item and the old Runs table item (TanStack Virtual, Bits UI).

0.9.0

Media

  • Tables: run.add_table(key, rows) logs per-example results, with audio and image cells (nb.Audio, nb.Image). The viewer compares each cell with the same row at the previous or first step, tinted green or orange with the change next to it, and rows glide to their new place as the step slider moves. Tables sort, filter (snr_in < 0 and noise = babble), show a trend per row and open a row's history and media. They export to HTML, and RunHandle.table() reads them back. The speech-denoise demo logs a table of 40 test files per eval, and its new listening-notes report shows it with audio notes.
  • PR curves: run.add_pr_curve(key, labels, predictions) counts true and false positives at evenly spaced thresholds, as TensorBoard does, and stores only the counts. Each run's curve shows with the figures, all runs are overlaid on one chart above them, and hovering shows the threshold and counts. nuboard import-tensorboard imports TensorBoard's PR curves, and the vit-ablations demo logs three of them.
  • Point clouds: run.add_point_cloud(key, points, colors=...) logs an (N, 3) cloud (lidar scans, mesh samples, 3D projections of embeddings). The Media page draws it in a view you can rotate, pan and zoom, colored by the logged colors or by height, with a step slider per run. It works in reports and in HTML exports. The vit-ablations demo logs eval embeddings that separate into classes over training.
  • Audio notes: select a section in the waveform player and add a note. Notes can be edited and deleted, are kept in the run's annotations.json, show in reports, and are listed under the clip in HTML exports, where clicking one plays its section. Run.audio_notes reads them from Python.

0.8.3

Runs

  • Property values must fit their type: a choice is picked from a menu, tags are toggled as colored pills, numbers, dates and links are checked, and the server and Run.set_field refuse values that do not fit. New options are added on purpose (New option…, + option) instead of by typing.
  • Removing a property removes its definition and its values; properties set without a definition can be removed too.
  • Run tags are colored pills under the run name instead of sitting in the config column.

Media

  • The key filter shows how many keys match and says so when none do.

Charts

  • Filter metrics (regex) above All metrics.
  • Points: a dot on each curve at the logged point under the cursor.
  • "robust y" is now ignore outliers, with a tooltip explaining it.

Other

  • About at the bottom of the side bar: version, copyright, license and a link to the documentation.

0.8.2

  • Exported reports show the NuBoard logo only as the browser tab icon; the default footer names NuBoard and its version instead.
  • The docs say NuBoard for the app and nuboard for the package, command and import.

0.8.1

Reports

  • Export as HTML asks for a theme (light, dark, or switchable with a button that follows the reader's system theme), the footer text (by default "Exported with NuBoard …") and whether to show the NuBoard logo in the footer and as the tab icon.
  • Charts saved from the Charts page are drawn live in the exported file from their data, so they follow its theme instead of showing a snapshot taken in another theme.

Runs

  • Add property picks the type from a menu instead of a text box.

0.8.0

Reports

  • Export as HTML: one self-contained file that opens offline, with images, audio and snapshots inlined, logged text and HTML included, and interactive Plotly and Vega-Lite figures whose libraries are embedded in the file (no CDN).
  • Sweeps and Timeline plots can be added to reports as snapshots, with an Open in Sweeps / Open in Timeline button that restores the view.

Sweeps

  • Grid for the scatter plots: plots per row and rows per page, with a pager.
  • Full screen for the parallel coordinates; a bookmark button adds it, or any scatter, to a report.

Timeline

  • Full screen, and a bookmark button that adds the timeline to a report.

Media

  • Select for report: tick several cards and add their items to a report in one go.
  • A chip per run with All and None, as in the Charts legend: click to hide a run, double-click or Alt-click to show only that run.
  • Logged HTML has output / code: the rendered page or its source with line numbers.

Other

  • The browser tab and the header say NuBoard.
  • Run details tabs are capitalized: Notes, Properties, All settings.
  • Fixes: embeds of the previous report no longer flash when switching reports.

0.7.0

Media page

  • Kind chips show or hide images, audio, figures, HTML and text (double-click one to show only that kind), and a regex filter narrows the keys.
  • Each key is a collapsible section with a card per run, and each card has its own step slider, as in TensorBoard. by run turns it around: a section per run with a card per key.
  • Step: latest, one step for every card, or best metric, which moves each run to its best step for a metric you pick (lowest or highest).
  • Grid: columns per row (fit, 1 to 8) and rows per page (1 to 6, or all), with a pager per section. Up to 48 runs.
  • Each item can be opened large, downloaded or added to a report.
  • The demo's vit-ablations project has more metrics (one per class), typed properties, a Markdown note and a sample report with images, a figure and a chart.

Reports

  • New Reports page: Markdown reports in <project>/reports/<name>.md, with a split editor and preview, autosave, rename, delete, print and Export .zip (the Markdown plus the embedded files).
  • ![caption](media:<run id>/<key>@<step>) embeds a logged image, audio clip, figure, HTML page or text. The Media page writes these lines for you.

Charts page

  • Grid for the small charts: charts per row (fit, 1 to 6) and rows per section (1 to 6, or all), with a pager per section.
  • Full screen for the focused chart and its inspector.
  • Add to a report: saves the focused chart as shown (runs, grouping, smoothing, zoom, hidden series). Reports draw it live and keep a PNG snapshot for exports and other Markdown viewers.

Notes and properties

  • Run notes are Markdown, edited with a preview in the run's Notes tab, and can embed logged items.
  • Typed properties (text, number, choice, tags, bool, date, link), defined per project in fields.json and set in the run's Properties tab. They are custom fields, so they show up as columns and in queries.

Config files

  • NuBoard no longer writes files/config.json. run.add_file(path) copies any file or folder (YAML, TOML, Hydra outputs, ...) into the run's files/, where the viewer shows and diffs it. The config passed to init is still in run.json.

Fixes

  • Charts kept their first width: a chart drawn before its container was laid out (or after a window resize) did not fill it.
  • Two runs with the same name no longer break the Charts page; their series are labelled with the run id's suffix.
  • Long run, metric and key names no longer overflow the Charts inspector, the summary cards (Best, Running, ...) and other panels; they are cut with an ellipsis and shown in full on hover.

0.6.0

TensorBoard import

  • nuboard import-tensorboard LOGDIR imports TensorBoard logs as runs: every folder of events.out.tfevents.* files under LOGDIR becomes a run, named after its path, with its parent path as the group. Scalars, images, audio, text and HParams hyperparameters (stored as the run's config) are imported with their original steps and wall times. Histograms, graphs and other plugins are skipped and counted. Logs from TensorFlow 1 and 2, PyTorch's SummaryWriter and tensorboardX are supported, and importing the same folder again skips the runs already imported.
  • The event files are decoded with the standard library, so the importer needs neither TensorFlow nor TensorBoard, and adds no dependency.
  • From Python: nuboard.tensorboard.import_logdir(logdir, project, dir=..., tags=...).
  • add_text, add_image, add_audio, add_figure and add_html take a keyword-only time= timestamp, like run.log.

Fixes

  • Queries can use the group field (where group = "fit/run1", group by group); before, group was only read as the start of group by.

0.5.0

Breaking: media methods are now add_*

  • run.log_text, log_image, log_audio, log_figure, log_html and log_artifact are renamed to add_text, add_image, add_audio, add_figure, add_html and add_artifact, with the same arguments. run.log({...}) for scalars keeps its name. The on-disk format does not change, so runs logged with older versions open as before.

Code style

  • Google-style docstrings throughout the package, checked by ruff's pydocstyle rules. Inline code in docstrings uses single backticks.
  • A blank line separates every if, for, while, with, try and match block from the statements around it. scripts/blank_lines.py applies the rule and a test checks it.

License

  • NuBoard is released under the MIT License (LICENSE, linked from the README and the docs).

0.4.3

  • Charts: All and None buttons next to the legend, and double-click (or Alt-click) on a legend chip to solo that series.
  • Save current view, Set field…, Add to collection and deleting a collection now ask in a dialog that follows the NuBoard theme instead of the browser's own prompt.

0.4.2

  • Custom fields are easier to find and edit: the + in Columns is disabled until a name is typed (before, it silently did nothing), adding a field closes the menu to show the new column, and the menu explains how to edit cells.
  • Set field… in the selection bar sets one field on all selected runs.
  • RunHandle.set_field(name, value) writes a custom field from Python.

0.4.1

  • Spectrogram options for the waveform player on the Media page: colormap (theme, viridis, magma, gray), frequency scale (linear, log, mel), FFT window (256 to 4096) and a switch to hide the spectrogram. They are remembered per browser and kept in the URL.

0.4.0

Audio

  • run.log_audio(key, audio, step, sample_rate=..., caption=...) logs a file path (WAV, FLAC, MP3, Ogg/Opus, M4A, WebM), encoded bytes, or a NumPy/torch array. Arrays are written as 16-bit WAV with the standard library; sample rate, channels and duration are stored with each clip.
  • The Media page plays audio with a waveform player (waveform and spectrogram, zoom, minimap, selection and loop, speed, full screen; the players of a key follow each other and only one plays at a time), ported from audioviewer and built on wavesurfer.js, or with the browser's simple player.
  • Audio is served with HTTP Range support, so players can seek in long clips.
  • The demo has a fourth project, speech-denoise: three models whose enhanced output gets cleaner over epochs, next to the noisy input and the clean target.

API

  • The documented import is now import nuboard as nb (it was nm). Nothing in the package changes; only the examples, docs and tests use the new alias.

0.3.1

  • Compare takes up to eight runs (it was limited to two). All settings shows a column per run and highlights rows where any run differs; file tabs diff any two of the compared runs.
  • The run details and compare panel now open above the runs table and scroll into view, instead of appearing below a long table where they were easy to miss.

0.3.0

Charts

  • X axis by any metric: plot every chart against another metric, such as train/tokens or epoch, from the new x by metric menu (x=metric:<name> in the URL). Points are matched by step, with the x metric interpolated where it was logged at other steps. Zoom maps back to a step range when the x metric only grows, so finer data still loads on zoom.

Media

  • Interactive figures: Plotly, Vega-Lite (Altair) and Bokeh figures render interactively on the Media page, each in a sandboxed iframe that follows the theme. Plotly and Vega are bundled and load on demand; Bokeh loads BokehJS for the figure's exact version from cdn.bokeh.org.
  • log_figure accepts Plotly figure dicts ({"data": [...], "layout": ...}) without Plotly installed.

Other

  • The demo logs a Vega-Lite per-class accuracy chart and a Plotly confusion matrix in the ViT ablations.
  • Built assets are served with Access-Control-Allow-Origin: * so the sandboxed figure frame can load them.

0.2.0

New pages

  • Sweeps: parallel coordinates over the parameters that vary between runs, with brushing on every axis, log axes for wide numeric ranges, a scatter per parameter, a best-runs table and a rank correlation per parameter.
  • Timeline: every run as a bar over wall-clock time, one row per run or packed to show concurrency, colored by status, project or group, with crash, alert and artifact markers and drag-to-zoom.
  • Artifacts: every artifact version with files, hashes and metadata, the runs that produced and used it, and a lineage graph.

Runs table

  • Health alert badges next to run names.
  • Columns menu: add any metric, config key or custom field as a column.
  • Editable custom fields, stored in annotations.json and usable in queries (verdict = keep) and as sweep parameters.
  • Collections: named sets of runs in <project>/collections/, listed in the workspace and queryable with collections has "name".

Health alerts

  • Non-finite values, loss and gradient spikes, and throughput drops are detected per metric and rank, stored in run.json at compaction and computed live for running runs. They appear in the runs table, under charts (with shaded regions) and on the timeline, and through RunHandle.alerts().
  • Crashed and killed runs are now compacted too, so their alerts are stored.

Other

  • New API endpoints: /api/alerts, /api/collections; /api/artifacts now returns producers, consumers and sizes.
  • Query fields alerts, collections and fields.<name>.
  • The demo gained a loss spike, an evaluation run that uses the checkpoint and writes a report artifact, two collections and custom fields.
  • The frontend approves its build scripts in pnpm-workspace.yaml, which pnpm 10 and 11 read (pnpm 11 ignores the pnpm field in package.json).
  • The documentation was reorganized into Installation, Quickstart, Guides and Reference.

The file format stays at version 1: all changes add keys or files, which older readers ignore.

0.1.0

First version: the log format, Python API, reader, Parquet compaction, analysis API, SQLite catalog, query language, and the viewer's Runs, Charts and Media pages.