---
library_name: transformers
base_model: Qwen/Qwen3-8B
tags:
- ronpo
- mnpo
- preference-optimization
---

# qwen3-8b-aaai27-flagship-dpo-s42

Research checkpoint for the RONPO AAAI revision experiments.

- Method: dpo
- Base model: `Qwen/Qwen3-8B`
- Seed: 42
- Local source at upload time: `/NHNHOME/WORKSPACE/26msit001_A/BASE/aipr_lab_sjkim_eval/revision_qwen3_8b/full_iter1/flagship_20260712/full/dpo/seed42/attempt1`
- Uploaded at UTC: 2026-07-12T17:21:45Z
- Run status metadata: `{"attempt": 1, "effective_batch_size": 16, "method": "dpo", "objective_protocol": "objective_protocol.json", "optimizer_steps": 900, "seed": 42, "split_manifest": "split_manifest.json", "stability_gate": "stability_gate.json", "status": "completed", "wandb_run_id": "f11bcc5bf0a6", "wandb_url": "https://wandb.ai/promotion-kim/mnpo/runs/f11bcc5bf0a6"}`

AAAI-27 P1 matched budget: 900 optimizer steps, effective batch 16; passed non-thinking and collapse stability gates; sealed-test metrics are not used for checkpoint selection.

Intended use: reproducibility and evaluation for the RONPO paper. This
checkpoint is not intended as a production assistant.