| license: other | |
| tags: [meta-reasoning, qwen3_5, condgen] | |
| base_model: HerrHruby/MR_midtrain_9B_v4 | |
| # MR_midtrain_9B_v4_condgen | |
| `HerrHruby/MR_midtrain_9B_v4` with the vision tower re-keyed from | |
| `model.language_model.visual.*` to `model.visual.*`. **Weights are byte-identical**; | |
| only 333 tensor NAMES change. | |
| ## Why | |
| verl's Megatron bridge (mbridge) resolves the tower by literal key name and fails | |
| on the released layout with | |
| ``` | |
| KeyError: 'model.visual.patch_embed.proj.weight' | |
| ``` | |
| transformers 5.5.4 and vLLM both load either layout — they resolve through the HF | |
| module tree — so the released repo is fine for inference and unusable for | |
| Megatron training. Both v3 artifacts use the flat layout, which is why v3 trained | |
| and v4 did not. This export reconciles against the working v3 export at **0 | |
| unmatched keys**. | |
| Use this repo for RL/Megatron training; either repo works for inference. | |
| Regenerate with `meta_reason_rl/e_rl/scripts/fix_v4_condgen_keys.py`. | |