FlaneAI's picture
Update README.md
11b3c1c verified
|
Raw History Blame Contribute Delete
2.77 kB
---
pretty_name: ResearchForge RL — Scientific Discovery Environments
language:
- en
license: mit
tags:
- rl-environment
- reinforcement-learning
- scientific-discovery
- multi-turn
- procedural-generation
- tool-use
- verifiable-rewards
size_categories:
- n<1K
---
# ResearchForge RL v0.5.0 — Hugging Face RL Environment Repository
**A multi-turn, verifiable scientific-discovery taskset and Python runtime for LLM agents.**
> **Compatibility:** This repository uses the Hub's `rl-environment` tag. It ships a working **native Python runtime**, not a verified OpenEnv, Verifiers, Harbor or NeMo Gym integration. We intentionally do not claim those framework tags. You can build adapters on top of `ResearchSession`.
## Install and run locally
```bash
pip install -e '.[dev]'
python scripts/smoke_test.py
pytest -q
```
```python
from researchforge.hub import load_tasks, ResearchSession
task = load_tasks('train')[0]
session = ResearchSession(task)
state = session.reset()
print(state['instructions'])
print(state['action_schema'])
result = session.step({'action': 'observe', 'n': 2})
print(result)
```
The model receives observations and JSON action descriptions. Repeat `session.step(action)` until `terminated` or `truncated`; terminal scores are computed by the simulator, not a language-model judge.
## Train / evaluate correctly
* Use `train` tasks for SFT, preference optimization and RL rollouts. Use `val` for selection; reserve `test` for final reporting.
* **Public seeds and source code are not a secure blind benchmark.** A malicious or source-aware agent could reconstruct worlds. For meaningful held-out evaluation, run the environment on a trusted machine, provide a **private** `evaluator_secret` in `ResearchSession`, keep it off the Hub, and use newly sampled private task seeds. The private evaluator must not expose world objects, source files or oracle details to the model.
* Do not publish private salts, oracle states, or answers. Do not train on test tasks.
* No LLM weights are trained by this repository. The example policies are hand-coded baselines, not neural training results.
* The simulator is research software, not evidence of real-world scientific discovery.
## Repository structure
- `data/{train,val,test}.jsonl`: task descriptors (not answers)
- `researchforge/hub.py`: `load_tasks`, `ResearchSession`, `rollout`
- `researchforge/env.py`: underlying step/reset/reward engine
- `researchforge/server.py`: isolated-process JSONL protocol (not a hardened security sandbox)
- `researchforge/training_data.py`: SFT / preference data utilities
- `researchforge/evaluation.py`: paired evaluation utilities
- `examples/`: agent and benchmark examples
- `tests/`: unit and integration tests
## License
MIT. See `LICENSE`.