| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen3.5-4B |
| pipeline_tag: text-generation |
| tags: |
| - qwen3.5 |
| - megatron-core |
| - sdpo |
| - reinforcement-learning |
| - math |
| - code |
| - search |
| --- |
| |
| # Qwen3.5-4B MathCodeSearch SDPO-E — Step 30 |
|
|
| This repository contains the complete raw Megatron Core distributed checkpoint |
| from the Qwen3.5-4B Math+Code+Search SDPO-E run at rollout step 30. |
|
|
| - Base model: [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) |
| - Training run: [Weights & Biases run `jzqlsb95`](https://wandb.ai/isaac_ghx-ipfipfipf/miles-sdpo/runs/jzqlsb95?nw=nwuser1004271927) |
| - Checkpoint iteration: `iter_0000029` (rollout step 30; rollout indices are zero-based) |
| - Format: Megatron Core `torch_dist` distributed checkpoint |
| - Tensor-parallel size used for training: 4 |
| - Total uploaded checkpoint size: 58,886,302,059 bytes (54.842 GiB) |
|
|
| ## Contents |
|
|
| The upload preserves the original checkpoint tree, including all 16 `.distcp` |
| shards, `.metadata`, `common.pt`, `metadata.json`, |
| `latest_checkpointed_iteration.txt`, and the rollout dataset state files for |
| steps 10, 20, and 30. |
|
|
| ## Loading |
|
|
| This is **not** a Transformers `from_pretrained` checkpoint. Resume or inspect |
| it with a compatible Megatron Core/MILES-SDPO environment and point the |
| checkpoint loader at the repository root. The original tensor-parallel layout |
| and the `torch_dist` checkpoint format should be retained when resuming. |
|
|
| This is an intermediate checkpoint; the run stopped after step 30 rather than |
| completing the planned 51 rollouts. |
|
|
