FineEnvs/multi-harness-rl

Space

Multi-Harness RL

18

31 commits

1 linked in READMEs

updated Oct 1, 2026

See the code

README

Multi-Harness RL

Read the article.

A research article built with research-article-template.

Source lives in FineEnvs under content/articles/multi-harness-rl/.

Quick start

cd app
npm install
npm run dev           # http://localhost:4321

Where the content lives

PathWhat
app/src/content/article.mdxFrontmatter and the chapter registry β€” the explicit import list is the running order
app/src/content/chapters/One .mdx per section
app/src/content/embeds/Standalone HTML/D3 visualizations, one file each
app/src/content/assets/image/Images
app/src/content/assets/data/Data files, served at /data/<name>
app/src/content/bibliography.bibReferences, cited as [@key]

Deploy

From the repo root, over the Hub HTTP endpoint (no git remote, no nested repo):

python3 tools/deploy.py content/articles/multi-harness-rl FineEnvs/multi-harness-rl

The Dockerfile and nginx config are included; this README is the Space card.

Review mode

The Space can take comments and suggested edits on the article, Google Docs style, without any of it showing on the published page.

  • Published (default). With REVIEW_MODE unset or off, the page never loads the review layer and every /api/review/ endpoint answers 404.
  • Review. With the Space variable REVIEW_MODE=on, open the article with ?review (https://fineenvs-multi-harness-rl.hf.space/?review) and sign in with Hugging Face. Select text to comment, suggest a deletion or suggest new wording, or use a figure's Comment button. Cards sit in the margin next to their text; the Comments panel lists open, resolved and detached threads. Only the Space owner and the usernames in REVIEWERS can read or write, and only the owner can accept or reject a suggestion. Exit (or ?review=off) goes back to the published view.

For an organization-owned Space, set REVIEW_OWNER to the maintainer's HF username. This keeps review moderation tied to a person rather than the organization name.

Threads are stored one JSON file each under /data/review-comments/threads/, on the private bucket AdithyaSK/multi-harness-rl-review, mounted at /data (Space settings β†’ Storage Buckets). To read them all at once:

uv run --with huggingface_hub python review/pull_comments.py   # writes review/.pulled/comments.md (git-ignored)

The script calls the owner-only /api/review/export endpoint with your local Hugging Face login; /api/review/import takes threads back in the same format.

The server is review/server.py (FastAPI, run next to nginx by entrypoint.sh); the page side is app/public/review/. For local testing, run the server with COMMENTS_DIR pointing at a folder and REVIEW_MODE=on; the dev server proxies /api/review and /oauth to it, and sign-in is mocked with your local Hugging Face login.

agents
docker
grpo
harbor
llm-training
openenv
reinforcement-learning
research-article-template
rl-environments

FineEnvs/multi-harness-rl

Space

Multi-Harness RL

18

31 commits

1 linked in READMEs

updated Oct 1, 2026

See the code

README

Multi-Harness RL

Read the article.

A research article built with research-article-template.

Source lives in FineEnvs under content/articles/multi-harness-rl/.

Quick start

cd app
npm install
npm run dev           # http://localhost:4321

Where the content lives

PathWhat
app/src/content/article.mdxFrontmatter and the chapter registry β€” the explicit import list is the running order
app/src/content/chapters/One .mdx per section
app/src/content/embeds/Standalone HTML/D3 visualizations, one file each
app/src/content/assets/image/Images
app/src/content/assets/data/Data files, served at /data/<name>
app/src/content/bibliography.bibReferences, cited as [@key]

Deploy

From the repo root, over the Hub HTTP endpoint (no git remote, no nested repo):

python3 tools/deploy.py content/articles/multi-harness-rl FineEnvs/multi-harness-rl

The Dockerfile and nginx config are included; this README is the Space card.

Review mode

The Space can take comments and suggested edits on the article, Google Docs style, without any of it showing on the published page.

  • Published (default). With REVIEW_MODE unset or off, the page never loads the review layer and every /api/review/ endpoint answers 404.
  • Review. With the Space variable REVIEW_MODE=on, open the article with ?review (https://fineenvs-multi-harness-rl.hf.space/?review) and sign in with Hugging Face. Select text to comment, suggest a deletion or suggest new wording, or use a figure's Comment button. Cards sit in the margin next to their text; the Comments panel lists open, resolved and detached threads. Only the Space owner and the usernames in REVIEWERS can read or write, and only the owner can accept or reject a suggestion. Exit (or ?review=off) goes back to the published view.

For an organization-owned Space, set REVIEW_OWNER to the maintainer's HF username. This keeps review moderation tied to a person rather than the organization name.

Threads are stored one JSON file each under /data/review-comments/threads/, on the private bucket AdithyaSK/multi-harness-rl-review, mounted at /data (Space settings β†’ Storage Buckets). To read them all at once:

uv run --with huggingface_hub python review/pull_comments.py   # writes review/.pulled/comments.md (git-ignored)

The script calls the owner-only /api/review/export endpoint with your local Hugging Face login; /api/review/import takes threads back in the same format.

The server is review/server.py (FastAPI, run next to nginx by entrypoint.sh); the page side is app/public/review/. For local testing, run the server with COMMENTS_DIR pointing at a folder and REVIEW_MODE=on; the dev server proxies /api/review and /oauth to it, and sign-in is mocked with your local Hugging Face login.

agents
docker
grpo
harbor
llm-training
openenv
reinforcement-learning
research-article-template
rl-environments