MR_midtrain_9B_v5_condgen

This is the final step-4128 checkpoint, not the best-validation checkpoint, of a six-epoch V5 supervised midtraining run from the stock Qwen3.5-9B weights. The model uses the Qwen3_5ForConditionalGeneration architecture and is published with the vision tower in the flat model.visual.* key layout expected by verl's Megatron bridge. The export changes only the names of 333 vision tensors from model.language_model.visual.*; their values are not modified. The vision tower is carried from the base model and was not a target of the text-only SFT run.

Training data and format

The training split contained 88,115 examples: 49,188 exploration (E), 29,450 meta-reasoning (MR), 6,477 final-answer (FA), and 3,000 termination traces. It was derived from HerrHruby/MR_midtrain_V5_sft with additional termination traces sampled from existing frontiers. The unchanged validation split had 258 examples. Training used the V4 verl/FSDP recipe with a 65,536-token maximum training length and the V5 attribution-aware tokenizer.

FA training targets retain the FA chain of thought through </think>, then emit <attribution>...</attribution> before the final answer or proof. The tokenizer preserves the V4 <direction>, </direction>, <summary>, and </summary> tokens (IDs 248077–248080) and adds <attribution> and </attribution> as atomic special tokens (IDs 248081–248082). The model's embedding size is unchanged.

The final checkpoint's validation loss was 0.642673. The lowest validation loss in the run was 0.642636 at step 2700, but that checkpoint is not the one published here. These losses are training diagnostics, not a downstream benchmark evaluation.

Compatibility

The flat vision-tower key layout matches the published MR_midtrain_9B_v4_condgen layout and is intended for verl/Megatron loading. It also retains the full conditional-generation architecture for Transformers and vLLM. For text-only vLLM serving, disable image and video inputs (for example with --limit-mm-per-prompt '{"image":0,"video":0}').

The export process first merges the final verl FSDP model shards into a Hugging Face safetensors checkpoint, then re-keys the 333 vision tensors. Optimizer and dataloader states are not part of this inference/RL-init export.