agentorch-rl-gemma26b-step680

RL checkpoint from the AgentOrch training project (gemma-26b), evaluated on the AgentOrch benchmark (Global View, 3,220 variants, frozen harness):

orchestration accuracy	final accuracy	error rate
0.2393	0.0935	0.4823

Training recipe (from training/results/training.csv, row rl-gemma26b-step680):

• SFT: v1 data, 2 epochs, 81920-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 680/992 (48h limit), data data_trl_v1, initialised from sft-gemma26b-v1-2ep

Base model: google/gemma-4-26B-A4B-it at revision 4d7ae4984b7db7de8f8457170b3f1a419ee76d52. 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-gemma26b-step680-global. See the AgentOrch repository for the harness, the training data and the full experiment record.