Qwen3.5-4B OPD — base, step 20
Research checkpoint after 20 completed OPD optimizer updates.
Zero-based Megatron iteration: 19. These OPD updates are
counted from this arm's initialization, not from the preceding SFT training.
Table with columns: Setting, Value| Setting | Value |
|---|
| Initialization | Qwen/Qwen3.5-4B-Base |
| Initialization revision | 1001bb4d826a52d1f399e183466143f4da7b741b |
| Teacher | Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 |
| Objective | Plain reverse-KL OPD; no additional reference KL |
| Prompt pool | Latest eight-dataset mixture, preserved source prompt order |
| Learning rate | Constant 1e-06 |
| Batch | 30 prompts × 4 responses |
| Generation / sequence limits | 32768 / 65536 tokens |
| Sampling | Temperature 1.0, seed 42 |
Training run: opd-q38t-base-lr1e6-20steps-save5-20261118-v1.
Teacher uses the archived dataset-specific system: Spec on harmful/difficult
prompts and no added Spec on benign prompts.
Inference weights and tokenizer/configuration assets only. Trained language
weights are converted from the committed model checkpoint and verified by exact
inverse round-trip. Vision/MTP tensors absent from the training checkpoint are
retained unchanged from this arm's pinned initialization. Optimizer/RNG state,
training data and credentials are excluded. No benchmark improvement is claimed.