Chapter 1 - 9B beyond-accuracy evaluator, LoRA adapters
LoRA adapters for Qwen3.5-9B (bf16). Each grpo_vNN was produced by SFT
followed by GRPO; each sft_vNN is the SFT stage alone, i.e. the starting point
of the corresponding GRPO run.
Only the final adapter of each run is uploaded. The intermediate
checkpoint-150/175/200 saves stay on the training machine -- they are training
trajectory, not results.
grpo_v22 and grpo_v23 differ only in SFT data (chapters 1-5, 4195 rows vs
chapters 1-7, 6955 rows); same hyperparameters, reward, and GRPO step budget.
Training code: private repo pigeonai-org/BeyondAccuracyEvaluation, branch
chapter1-9b-grpo-v22-v23. Evaluation results and datasets: private dataset
seleixi/beyondacc-chapter1-backup.
Prompt contract differs across versions
v17-v21 were trained to emit seven aspect scores plus mean; v22 onward
emit only overall, nominal_label, rationale. Feeding a v22+ adapter a v21
prompt asks it for fields it was never trained to produce. Use
rewrite_eval_prompt.py from the training branch to convert an eval file.
- Downloads last month
- -