| license: other | |
| base_model: Qwen/Qwen2.5-1.5B-Instruct | |
| tags: [dpo, lora, self-training, label-quality] | |
| # Self-Training Loop — Phase 2 A'' three-arm ablation adapters | |
| Six LoRA adapters (Qwen2.5-1.5B-Instruct, r16, lr 1e-5, beta 0.1, 2 epochs, seeds 22/23): | |
| - `D_full_emf1_seed{22,23}` — 10,000 preference pairs, original EM/F1 directions | |
| - `D_full_corrected_seed{22,23}` — same 10,000 pairs with 375 judge-verified hurts-direction flips | |
| - `D_verified_seed{22,23}` — 3,185 semantically verified pairs (hybrid accept pool) | |
| Companion dataset repo: `self-training-loop-phase2-arms` (training arms, eval chain, | |
| human adjudication records). Key result: direction repair at 3.75% flip rate produces | |
| no detectable behavioral difference (two seeds, reference-based and blind pairwise | |
| LLM-judge evaluation) — see the thesis repo Phase 2 execution status doc §15.22. | |