pagent-hintflow-dpo-v2-merged
Merged full-weight Qwen3-4B HintFlow orchestrator after offline DPO on tree-exported preference pairs (v2 trees).
Training snapshot
- Base:
Qwen/Qwen3-4B
- Data:
hintflow_trees_v2 → 13565 DPO pairs (plan 466 / review 13099), τ=0.05
- Train: 3 epochs, LoRA r=16, β=0.1, lr=5e-6, world_size=3, effective_batch=9
- Val pair-acc (preference ranking, not EM): e1 50.2% / e2 50.2% / e3 50.9%
Files
This repo is the merged checkpoint used for vLLM serve (served-model-name typically qwen3-4b).
Companion LoRA adapter (if uploaded): ivaning0919/pagent-hintflow-dpo-v2-adapter.
Note
Backup / archival upload from the PAgent project. Preference-pair accuracy stayed near chance; use downstream HintFlow 128 EM for real evaluation.
