agentorch-rl-gemma31b-final

RL checkpoint from the AgentOrch training project (gemma-31b), evaluated on the AgentOrch benchmark (Global View, 3,220 variants, frozen harness):

orchestration accuracy	final accuracy	error rate
0.191	0.0922	0.5655

Training recipe (from training/results/training.csv, row rl-gemma31b-final):

• SFT: v1 data, 1 epochs, 65536-token budget, LoRA r16/a32, lr 1e-4, schedule cosine
• RL: GRPO (TRL GRPOTrainer, LoRA r16/a32, colocate vLLM TP8), reward Tier-1 symbolic match, lr 1e-5, KL beta 0.001, group 8, 8 prompts/step, steps 40 (131 prompts, completion cap 3072), data data_trl_v1_31b (prompts <=210k chars, 500 rows), initialised from sft-gemma31b-v1-1ep

Base model: google/gemma-4-31B-it at revision 842da3794eaa0b77d5f08bae87a17459d91ff475. Merged weights (LoRA folded into the base); load with transformers or vLLM as a normal checkpoint. Use enable_thinking=False in the chat template, as at training and evaluation time.

Evaluation run: rl-gemma31b-final-global. See the AgentOrch repository for the harness, the training data and the full experiment record.