Tensorboard reader

Skill jacobjennings/tensorboard-reader-skill/skills/tensorboard-reader

Read and summarize TensorBoard event logs (.tfevents) token-efficiently. Use when inspecting a training/eval run's logs — scalars (loss, lr, metrics), text samples, histograms, hparams, or images/audio — without dumping raw data into context. Start with a tiny overview, then drill down only where needed.From its SKILL.md

Install

npx -y skills add jacobjennings/tensorboard-reader-skill --skill tensorboard-reader

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its file declares

Copied from the file, not written here

The file declares its own license as Apache-2.0. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

4.0 KB, 993 tokens by cl100k_base, as published. Nobody here has run it

TensorBoard Reader

Inspect TensorBoard logs with progressive detail: a cheap overview first, drilling down only where it matters. The reader never dumps full series or binary blobs unless you explicitly ask, so reading a run costs a few hundred tokens instead of tens of thousands.

All commands go through one script: scripts/tbread.py.

Quick start

uv run scripts/tbread.py overview <LOGDIR>

<LOGDIR> is a directory containing TensorBoard runs. Run discovery is recursive — it finds nested run subdirs (e.g. runs/Jun12_*) and merges the many event files a single dir may contain (multi-process training) into one logical run. overview is the default, so uv run scripts/tbread.py <LOGDIR> works too.

Progressive-detail workflow (always overview first)

Tier	Command	What it costs / gives
0	`overview LOGDIR`	Tiny map: runs → tags grouped by type, with point counts + step/wall ranges. Never prints values.
1	`scalars LOGDIR --tag T`	Default `--stats`: count/min/max/first/last/mean/trend. Add `--series` for a downsampled curve.
1	`text LOGDIR --tag T`	Newest-step text, truncated to `--max-chars` (default 2000).
1	`hist LOGDIR --tag T`	Histogram bucket summary at a step (coarse bins, not raw arrays).
1	`hparams LOGDIR`	Compact hparam table across runs.
1	`info LOGDIR --tag T`	Shape/dtype/count/bytes for any tag — the cheap way to size an image/audio/mesh tag before extracting it.
2	`export LOGDIR --tag T --out DIR`	Decodes bytes to files on disk (.png/.wav/.csv/.txt/...). The only command that writes bytes.

Rule of thumb: run overview, decide which tags matter, then pull just those with scalars --stats. Only reach for --series, text, or export when stats aren't enough.

Common flags

--tag PAT — glob over tag names; repeatable (--tag 'train/*' --tag eval/loss).
--run PAT — restrict to run(s) by name or basename glob; repeatable. Cheaper too (only matching runs are loaded).
--json — machine-readable output for every command.
scalars: --series (downsampled curve), --points N (cap, default 50, min/max-preserving so spikes survive), --uniform (plain stride), --full (every raw point — opt-in), --all (all scalar tags).
text: --step N (default newest), --all-steps, --max-chars N.
hist: --step N (default latest), --bins N (default 8).
export: --out DIR (required), --all (every step; default latest only).

Output contract

Binary/heavy types (images, audio, mesh, graph) are never inlined. overview and info report counts/shape/bytes; use export to get the actual files.
overview collapses any type's tag list past --max-tags (default 40) to (+N more).
Notes like duplicate steps merged or non-monotonic steps are surfaced so you don't misread merged multi-process logs.

Runtime & fallback

Preferred: uv run scripts/tbread.py … — uv reads the script's inline dependency (tensorboard) and runs it in an isolated, cached env. No global install, no tensorflow needed.
If uv is absent: python3 scripts/tbread.py … in an env that has tensorboard. On a missing import the script exits non-zero with the exact recovery command (it never auto-installs). scripts/requirements.txt pins the one dependency.

References

references/data-types.md — what each TensorBoard data type shows at overview vs drill-down vs export.
references/tensorboard-api.md — how the reader works and the TensorBoard API gotchas it handles.
references/examples.md — a worked HuggingFace-Trainer walkthrough.