msr-spare-1/spare-fixed-corpus-qwen3-30b-6skill-medium
SPARE fixed corpus — Qwen3-30B-A3B-Instruct-2507, 6 skills, medium 2,400 validated multi-turn environments (400 per skill) generated offline by base Qwen3-30B-A3B-Instruct-2507. This is the corpus consumed by the 0622 actor-only fixed-corpus ablation arm. 2400 validated environments. Skill Games Causal Inference 400 Logical Deduction 400 Mathematical Reasoning 400 Optimization 400 Pattern Recognition 400 Spatial Reasoning 400 Each game exposes the SPARE… See the full description on the dataset page: https://huggingface.co/datasets/msr-spare-1/spare-fixed-corpus-qwen3-30b-6skill-medium.
SPARE fixed corpus — Qwen3-30B-A3B-Instruct-2507, 6 skills, medium
2,400 validated multi-turn environments (400 per skill) generated offline by base Qwen3-30B-A3B-Instruct-2507. This is the corpus consumed by the 0622 actor-only fixed-corpus ablation arm.
2400 validated environments.
Each game exposes the SPARE contract (reset(seed=None), step(action) -> (obs, reward, terminated, truncated, info)). Load via the project loader, which injects the stdlib names and base classes generated games rely on:
from spare.core.envs.synthetic_game_env import make_synthetic_env
env = make_synthetic_env("game_000000_mathematical_reasoning.py")