Rollout Result Viewer

A lightweight web UI for browsing the per-step JSONL files that Relax writes during training and evaluation rollouts.

What it shows

Relax always writes a compact per-rollout summary as long as --save is set:

Subdir	Source	Written by
<save>/rollout_result/train/{rollout_id}.jsonl	training rollouts	save_rollout_result_jsonl
<save>/rollout_result/eval/{rollout_id}.jsonl	eval rollouts	save_eval_summary_jsonl

Each line is one sample: prompt, response, reward, response_length, total_length, status, group_index, and (for eval) dataset. Multimodal inputs are summarized rather than dumped.

Launch

python -m relax.entrypoints.visualize <save>/rollout_result --port 8080
# Then open http://localhost:8080

The data directory is a required positional argument. python -m relax.utils.visualize ... works too and takes the exact same arguments — the entrypoint is just a thin wrapper kept next to train.py for consistency with other Relax entrypoints.

The server auto-discovers train/ and eval/ under the data dir. If both exist, a [ train | eval ] toggle appears at the top of the page. If neither exists, the data dir itself is treated as a flat bucket — you can also point it directly at a folder of {step}.jsonl files.

Common flags

Flag	Default	Purpose
DATA_DIR (positional)	required	The directory to serve.
--port	8080	HTTP port.
--host	0.0.0.0	Bind address.
--cache-memory	4096 (MB)	Total RAM the in-process LRU may use. Bump it if individual train/*.jsonl files are large.
--cache-entries	20	Max number of files held in cache.
--base-path	""	URL prefix when serving behind a reverse proxy, e.g. --base-path /absproxy/8080.

What the page does

• Step dropdown — every {step}.jsonl in the active subdir, sorted by step number.
• Sample navigation — ← Previous / Next →, plus arrow-key shortcuts and a floating mini-nav in the bottom-right corner.
• Sample Info card — scalar fields (rollout_id, sample_index, reward, response_length, total_length, status, group_index, dataset).
• Prompt / Response / Label sections — pretty-printed with chat-template, tool-call and <think> highlighting. Long text scrolls in place; default scroll position is the end so you can see what the model just produced.
• Sort controls — re-order samples within a step by sample_index, reward, or response_length (ascending/descending).
• Cache controls — GET /api/cache/stats, POST /api/cache/clear.

Terminal UI (--tui)

For SSH sessions where opening a browser is inconvenient, the same command can render a textual-based terminal viewer:

python -m relax.entrypoints.visualize <save>/rollout_result --tui

Requires the optional packages textual and rich (pip install textual rich). Not in requirements.txt — web users have zero extra dependencies.

Flag	Default	Purpose
--tui	off	Use the terminal viewer instead of the web viewer.
--mask-str	regex for <|image_pad|>, <|imgpad|>, <|audio_comp_pad|>	Substrings replaced with * to keep multimodal prompts readable. Pass --mask-str "" to disable.

The TUI loads {step}.jsonl files from a single flat directory. When the data dir contains both train/ and eval/, the TUI defaults to train/; point --data-dir at <save>/rollout_result/eval to inspect eval instead.

Key bindings

Keys	Action
n / p	Next / previous sample
N / P	Next / previous step
f then type, enter	Find; enter again to jump to next match
esc	Clear search
s	Switch between plain-text and table rendering
r	Refresh current view
j / k	Page down / up
h / l	Page left / right
g / G	Top / bottom
tab / ← / →	Move focus between widgets

The left sidebar has dropdowns for step, sample, dataset (when eval data is loaded), and sort mode (reward asc/desc, response_length asc/desc), plus a per-field show/hide list.

Operational notes

• Intended for local or trusted-internal-network use. CORS is wide open and there is no authentication — same posture as the rest of the dump-inspection tooling.
• The server caches parsed JSONL files in process memory. For very large train dumps, raise --cache-memory or restart the process to free RAM.
• The viewer trusts on-disk JSON; malformed lines are skipped with a warning.