Datasets:

FlaneAI
/

Scientific-Research-RL

task_id stringlengths 14 14	domain stringclasses 4 values	difficulty int64 1 3	seed int64 10k 10.3k	split stringclasses 1 value	budget int64 30 30	max_steps int64 32 32	noise_sd float64 0.15 0.15
rf-train-00000	law	1	10,000	train	30	32	0.15
rf-train-00001	causal	1	10,001	train	30	32	0.15
rf-train-00002	rule	1	10,002	train	30	32	0.15
rf-train-00003	hypothesis	1	10,003	train	30	32	0.15
rf-train-00004	law	2	10,004	train	30	32	0.15
rf-train-00005	causal	2	10,005	train	30	32	0.15
rf-train-00006	rule	2	10,006	train	30	32	0.15
rf-train-00007	hypothesis	2	10,007	train	30	32	0.15
rf-train-00008	law	3	10,008	train	30	32	0.15
rf-train-00009	causal	3	10,009	train	30	32	0.15
rf-train-00010	rule	3	10,010	train	30	32	0.15
rf-train-00011	hypothesis	3	10,011	train	30	32	0.15
rf-train-00012	law	1	10,012	train	30	32	0.15
rf-train-00013	causal	1	10,013	train	30	32	0.15
rf-train-00014	rule	1	10,014	train	30	32	0.15
rf-train-00015	hypothesis	1	10,015	train	30	32	0.15
rf-train-00016	law	2	10,016	train	30	32	0.15
rf-train-00017	causal	2	10,017	train	30	32	0.15
rf-train-00018	rule	2	10,018	train	30	32	0.15
rf-train-00019	hypothesis	2	10,019	train	30	32	0.15
rf-train-00020	law	3	10,020	train	30	32	0.15
rf-train-00021	causal	3	10,021	train	30	32	0.15
rf-train-00022	rule	3	10,022	train	30	32	0.15
rf-train-00023	hypothesis	3	10,023	train	30	32	0.15
rf-train-00024	law	1	10,024	train	30	32	0.15
rf-train-00025	causal	1	10,025	train	30	32	0.15
rf-train-00026	rule	1	10,026	train	30	32	0.15
rf-train-00027	hypothesis	1	10,027	train	30	32	0.15
rf-train-00028	law	2	10,028	train	30	32	0.15
rf-train-00029	causal	2	10,029	train	30	32	0.15
rf-train-00030	rule	2	10,030	train	30	32	0.15
rf-train-00031	hypothesis	2	10,031	train	30	32	0.15
rf-train-00032	law	3	10,032	train	30	32	0.15
rf-train-00033	causal	3	10,033	train	30	32	0.15
rf-train-00034	rule	3	10,034	train	30	32	0.15
rf-train-00035	hypothesis	3	10,035	train	30	32	0.15
rf-train-00036	law	1	10,036	train	30	32	0.15
rf-train-00037	causal	1	10,037	train	30	32	0.15
rf-train-00038	rule	1	10,038	train	30	32	0.15
rf-train-00039	hypothesis	1	10,039	train	30	32	0.15
rf-train-00040	law	2	10,040	train	30	32	0.15
rf-train-00041	causal	2	10,041	train	30	32	0.15
rf-train-00042	rule	2	10,042	train	30	32	0.15
rf-train-00043	hypothesis	2	10,043	train	30	32	0.15
rf-train-00044	law	3	10,044	train	30	32	0.15
rf-train-00045	causal	3	10,045	train	30	32	0.15
rf-train-00046	rule	3	10,046	train	30	32	0.15
rf-train-00047	hypothesis	3	10,047	train	30	32	0.15
rf-train-00048	law	1	10,048	train	30	32	0.15
rf-train-00049	causal	1	10,049	train	30	32	0.15
rf-train-00050	rule	1	10,050	train	30	32	0.15
rf-train-00051	hypothesis	1	10,051	train	30	32	0.15
rf-train-00052	law	2	10,052	train	30	32	0.15
rf-train-00053	causal	2	10,053	train	30	32	0.15
rf-train-00054	rule	2	10,054	train	30	32	0.15
rf-train-00055	hypothesis	2	10,055	train	30	32	0.15
rf-train-00056	law	3	10,056	train	30	32	0.15
rf-train-00057	causal	3	10,057	train	30	32	0.15
rf-train-00058	rule	3	10,058	train	30	32	0.15
rf-train-00059	hypothesis	3	10,059	train	30	32	0.15
rf-train-00060	law	1	10,060	train	30	32	0.15
rf-train-00061	causal	1	10,061	train	30	32	0.15
rf-train-00062	rule	1	10,062	train	30	32	0.15
rf-train-00063	hypothesis	1	10,063	train	30	32	0.15
rf-train-00064	law	2	10,064	train	30	32	0.15
rf-train-00065	causal	2	10,065	train	30	32	0.15
rf-train-00066	rule	2	10,066	train	30	32	0.15
rf-train-00067	hypothesis	2	10,067	train	30	32	0.15
rf-train-00068	law	3	10,068	train	30	32	0.15
rf-train-00069	causal	3	10,069	train	30	32	0.15
rf-train-00070	rule	3	10,070	train	30	32	0.15
rf-train-00071	hypothesis	3	10,071	train	30	32	0.15
rf-train-00072	law	1	10,072	train	30	32	0.15
rf-train-00073	causal	1	10,073	train	30	32	0.15
rf-train-00074	rule	1	10,074	train	30	32	0.15
rf-train-00075	hypothesis	1	10,075	train	30	32	0.15
rf-train-00076	law	2	10,076	train	30	32	0.15
rf-train-00077	causal	2	10,077	train	30	32	0.15
rf-train-00078	rule	2	10,078	train	30	32	0.15
rf-train-00079	hypothesis	2	10,079	train	30	32	0.15
rf-train-00080	law	3	10,080	train	30	32	0.15
rf-train-00081	causal	3	10,081	train	30	32	0.15
rf-train-00082	rule	3	10,082	train	30	32	0.15
rf-train-00083	hypothesis	3	10,083	train	30	32	0.15
rf-train-00084	law	1	10,084	train	30	32	0.15
rf-train-00085	causal	1	10,085	train	30	32	0.15
rf-train-00086	rule	1	10,086	train	30	32	0.15
rf-train-00087	hypothesis	1	10,087	train	30	32	0.15
rf-train-00088	law	2	10,088	train	30	32	0.15
rf-train-00089	causal	2	10,089	train	30	32	0.15
rf-train-00090	rule	2	10,090	train	30	32	0.15
rf-train-00091	hypothesis	2	10,091	train	30	32	0.15
rf-train-00092	law	3	10,092	train	30	32	0.15
rf-train-00093	causal	3	10,093	train	30	32	0.15
rf-train-00094	rule	3	10,094	train	30	32	0.15
rf-train-00095	hypothesis	3	10,095	train	30	32	0.15
rf-train-00096	law	1	10,096	train	30	32	0.15
rf-train-00097	causal	1	10,097	train	30	32	0.15
rf-train-00098	rule	1	10,098	train	30	32	0.15
rf-train-00099	hypothesis	1	10,099	train	30	32	0.15

End of preview. Expand in Data Studio

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

pip install -e '.[dev]'
python scripts/smoke_test.py
pytest -q

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.

Downloads last month: 58