LegoEnv checkpoints

Checkpoints trained on environments built by LegoEnv (LegoEnv: Self-Evolving Agentic Environments via Modular Recombination, arXiv: TODO). Code: https://github.com/JazyJiang/LegoEnv · Data: https://huggingface.co/datasets/JazySong/LegoEnv

Folder	Base	Training	Corpus
qwen3.5-27b-sft/	Qwen3.5-27B	SFT, 3 epochs (step 489)	round-1 corpus plus the 3k draw, 5,229 trajectories
qwen3.6-27b-sft/	Qwen3.6-27B	SFT, 3 epochs (step 570)	round-2 corpus, 6,110 trajectories
qwen3.8-27b-grpo/	Qwen3.8-27B	GRPO, step 200, binary verifier reward	1,516 environments

All three are full-parameter checkpoints in safetensors, with the base model's tokenizer and chat template. Training recipe: SFT with global batch 32, peak learning rate 1e-5, cosine decay, 64K sequence length, teacher reasoning kept across turns, assistant turns that received a 4xx tool response excluded from the loss; GRPO from the unmodified base with group size 8, temperature 1.0, 2 epochs over 1,516 prompts, reward = the environment's programmatic verifier (no reward model, no judge).

These are the checkpoints reported in the paper's Table 1.

Evaluation

Protocol: AutomationBench 1.0.6, api toolset, 50 steps, one rollout per task, reasoning of earlier turns replayed; MCP-Mark 127 tasks; BFCL v4 official weighted average.

Checkpoint	AutomationBench	MCP-Mark	BFCL v4
Qwen3.5-27B base	3.2	44.9	67.6
qwen3.5-27b-sft	23.2	60.0	67.6
Qwen3.6-27B base	16.0	42.5	63.0
qwen3.6-27b-sft	32.0	68.5	72.4
Qwen3.8-27B base	37.6	73.2	71.0
qwen3.8-27b-grpo	40.3	77.6	72.0

Numbers are the paper's Table 1 (AutomationBench: mean of one to four evaluations of 600 tasks; MCP-Mark and BFCL v4: mean of four evaluations).

Usage

Standard transformers loading; use the bundled chat template and pass tools in the OpenAI function-calling format. The models expect the reasoning of earlier turns to be kept in the conversation (reasoning_content on assistant messages), as during training.

License

The checkpoints inherit the Apache-2.0 license of their Qwen3 base models. Training data and code are licensed separately (see the dataset and code repositories).