Training provenance
- Run:
openmobile2-9b-think-history-fromstep2-20261106
- Exported checkpoint:
global_step_71/actor/huggingface
- Method: Anchor-GRPO, with terminal goal-success reward, process-reward coefficient
anchor_lambda = 3.0, action advantage coefficient tool_advantage_kappa = 0.25, and format-error advantage penalty coefficient format_advantage_coef = 1.0.
- Policy loss fixed-length normalization:
1024.
This checkpoint uses nonzero process reward and action weighting. Evaluation results are not asserted in this model card.
model-provenance.json records the SHA-256 hashes and sizes from the checkpoint completion manifest. Optimizer state, distributed training shards, task traces, and runtime credentials are not included in this inference-model export.
Use a Transformers release supporting Qwen3_5ForConditionalGeneration, and preserve the included processor and chat template when serving the model.