Spaces:
Sleeping
Sleeping
Commit History
-
Add interactive frontend UI
Made-with: Cursorb0c701c -
Update runtime requirements
Made-with: Cursor5c277bb -
Merge origin/main into main
Made-with: Cursorcbbb3c4 -
fix: add root status endpoint
Return basic environment status at / so Hugging Face Space probes and browser visits do not show 404 while preserving OpenEnv API routes. Made-with: Cursor808bf4d -
docs: trim readme to submission essentials
Keep the Hugging Face metadata while removing quick start, API reference, and citation sections from the public README. Made-with: Cursor44b2880 -
docs: restore hugging face space metadata
Add valid Docker Space YAML metadata so Hugging Face can parse the README card configuration. Made-with: Cursor228f4a9 -
Update README.md by removing unnecessary metadata
SiddhUnverified signatureRemoved metadata and author information from README.63b4bb4 -
Remove author submission line from BLOG.md
SiddhUnverified signatureRemoved the submission line from the blog post.7c4f887 -
docs: replace private wandb link with exported logs
Add local training output, completion samples, scalar history, and charts while updating README and blog attribution to submission by the authors. Made-with: Cursoreaf43b6 -
docs: remove internal checklist from readme
Keep the public README focused on submission materials, results, usage, and citations. Made-with: Cursor3590d38 -
docs: add separate mini blog writeup
Move the narrative writeup into BLOG.md and keep the README focused on submission links, results, and usage. Made-with: Cursor86ab1d8 -
docs: rewrite readme as mini blog
Lead with a judge-facing story that explains the problem, environment, training result, and why epistemic robustness matters. Made-with: Cursor1388ca9 -
docs: add final submission readme and training logs
Document the environment, training evidence, model links, authors, plots, and remaining finale checklist for hackathon submission review. Made-with: Cursor6a72222 -
fix: load model snapshots from writable cache
Download base and adapter snapshots into the Space cache before loading so Transformers does not probe remote repos through /.cache. Made-with: Cursorb3382c5 -
fix: load peft base model with explicit cache
Load the LoRA base model and adapter separately so both Hugging Face downloads use the Space writable cache instead of falling back to /.cache. Made-with: Cursor40bdaf7 -
fix: pass explicit cache dir to model loader
Use the writable Space cache path directly for tokenizer and PEFT model downloads so Hugging Face libraries cannot fall back to /.cache. Made-with: Cursor3ca7ae0 -
fix: set model cache dirs before torch imports
Configure writable Hugging Face cache paths before model libraries initialize so lazy adapter loading works inside the Space runtime. Made-with: Cursor264201d -
feat: add optional trained model inference to env space
Add lazy LoRA adapter inference endpoints and GPU model dependencies while preserving the OpenEnv reset/step server routes. Made-with: Cursor46c2f98 -
fix: reduce grpo training runtime
Use a smaller hackathon training run with fewer episodes, generations, and completion tokens so the A100 Space can finish in a practical window. Made-with: Cursor842caac -
fix: parse chat completions in rewards
Normalize TRL chat completion outputs before applying JSON reward parsing so GRPO can score conversational generations. Made-with: Cursor6f3d9d6 -
fix: emit prompt column for grpo trainer
Current TRL GRPOTrainer reads dataset rows from the prompt key, so expose chat prompts there while preserving reward metadata columns. Made-with: Cursor6ceec85 -
fix: add grpo warnings compatibility attr
Initialize warnings_issued on the PEFT model so current TRL GRPOTrainer can attach token-estimation warnings without failing on Qwen LoRA wrappers. Made-with: Cursor633d604 -
fix: choose writable cache root at space startup
Avoid build-time root-owned cache directories by selecting a verified writable runtime root before importing Gradio, Transformers, or Unsloth. Made-with: Cursor3125dc1 -
fix: use writable runtime dirs in training space
Route Hugging Face, Transformers, WandB, Matplotlib, and checkpoint outputs to /tmp so model loading and training work on read-only HF Space filesystems. Made-with: Cursoreafb471 -
fix: patch torchao dtype imports for unsloth
Allow the training Space to import Unsloth with stable CUDA PyTorch by shimming optional TorchAO dtype probes before Transformers loads. Made-with: Cursora262689 -
fix: shim torch._pytree.register_constant for torchao compat
torchao uses register_constant which doesn't exist in stable torch. Patch it as a no-op before importing unsloth — safe because we use bitsandbytes/LoRA, not torchao quantization. Reverts to stable cu121. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>98317c2 -
fix: use PyTorch nightly for torchao compat
Latest torchao (pulled in by transformers) needs torch.int1 AND register_constant — the former is in 2.6 but the latter only in nightly 2.7. Switch to --pre nightly/cu121 index. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>a73c9e7 -
fix: switch to cu124 wheel index for torch 2.6.0
torch 2.6.0 is not available for cu121 (max 2.5.1). Switch to cu124 wheel index — A100 driver supports CUDA 12.4, torch.int1 present. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>d98c86c -
fix: pin torch==2.6.0 to satisfy torchao torch.int1 requirement
torchao (pulled in by latest transformers) uses torch.int1 which was added in PyTorch 2.6.0. cu121 wheels available for 2.6.0. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>b42845a -
fix: verbose per-step logs and BaseException catch in _run_training
Each slow phase (unsloth import, TRL import, training start) now logs explicitly so the UI shows progress. Exception catch broadened to BaseException with full traceback so nothing is swallowed silently. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>6fc6438 -
fix: capture transformers/TRL/unsloth logs in Gradio UI
TRL trainer uses Python logging module, not print — so _Tee on sys.stdout/stderr missed all epoch/step/loss output. Added _LogHandler attached to transformers, trl, unsloth, peft, accelerate loggers so every training log line appears in the UI. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>0b6be50 -
fix: add unsloth_zoo as explicit dependency before unsloth
unsloth requires unsloth_zoo but doesn't pull it automatically from git install. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>6f1b59d -
fix: install unsloth in Docker image at build time
Moving unsloth install from click-time to build time. pip install unsloth doesn't require GPU — only running the model does. Removes all the runtime install logic from space_runner.py. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>dee46e3 -
fix: chunk-read pip output splitting on \r and \n
pip progress bars use \r not \n — line-by-line readline blocks forever waiting for a newline that never comes. Reads 512-byte chunks instead and splits on both, so every progress update appears immediately. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>7ac7eb8 -
feat: stream pip install output live to UI logs
Use Popen + line-by-line stdout read so every download/install line appears in the Gradio log box in real time instead of buffering to end. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>d79940c -
fix: install unsloth to /app/pkgs to bypass HF Space permission issue
HF Spaces run as a non-root user with HOME=/ so pip fails writing to /.local. Fix: use --target /app/pkgs (always writable) and add to sys.path after install. Also detect torch version at runtime to pick the correct unsloth git extra (torch251 vs torch240). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>722cd66 -
fix: robust unsloth install — PyPI first, git fallback, log errors
colab-new extra fails on A100 docker environment. Now tries pip install unsloth (PyPI) first, then cu121-ampere-torch240 git variant. Captures and logs stderr so failures are visible in UI. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>fa5785a -
feat: add A100 training Space with manual-trigger Gradio UI
Dockerfile.train uses python:3.11-slim + CUDA PyTorch wheel (avoids 7GB pytorch/pytorch base). space_runner.py installs unsloth at click-time on the live GPU rather than at Docker build time. Auto-refresh via gr.Timer instead of demo.load(every=) for Gradio 4.x compatibility. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>77d8bcf
-
feat: optimize Space — dockerignore, dotenv secrets, multi-worker uvicorn
0eb0abc -
fix: use lean server requirements in Dockerfile (exclude unsloth/torch)
8c7e984 -
chore: add HF Space frontmatter and README draft
40b1abb -
feat: bulletproof GRPO training script + Colab notebook
67601e4 -
feat: rewrite env to be fully openenv-core compliant
7a8a0f0 -
chore: update baseline results with real facts (450 QA pairs)
4e71c52 -
feat: implement data pipeline
Add fact loading, corruption generation, and document synthesis for context corruption episodes, plus UTF-8 fact loading for Windows smoke tests. Made-with: Cursord661639 -
feat: baseline eval and GRPO training script
5f54992 -
Updated .gitignore
ef79a5f -
chore: add CLAUDE.md, README, update gitignore
c12488d -
feat: implement actions, reward, env, server
18fd5bc -
chore: pin dependencies
d09377a