--- library_name: transformers base_model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B tags: [direct-opd, direct-opd-phase2, sft, math] --- # Direct-OPD pilot — SFT-shifted base teacher (Phase 2) `deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B` @ `ad9f0ae0864d7fbcd1cd905e3c6c5b069cc8b562` supervised-fine-tuned for **one pass** over the 6400 pre-registered `sft_train` rows of `cmpatino/direct-opd-sft-deepmath-pilot-data` @ `22625ae5db434947195bf862c429cd94504a4809`. This is an experiment artifact of the Direct-OPD SFT-vs-RL policy-shift pilot: it is the *SFT-shifted* teacher whose divergence from the base teacher Phase 3/4 measures. It is not intended or evaluated as a general-purpose model. ## Configuration (pre-registered) | | | |---|---| | steps / global batch | 100 x 64 = 6400 examples (6400 rows available) | | optimizer | AdamW lr 5e-06, betas (0.9, 0.95), eps 1e-08, wd 0.1 (no decay on 1-D params) | | schedule | linear decay, 10 warmup steps, grad-norm clip 1.0 | | precision | fp32 master weights, autocast bf16 compute, **bf16** checkpoints | | max sequence | 3072 tokens (prompt <= 1024, target + EOS <= 2048) | | supervised tokens | 10,950,054 | | seed | 42 (data order `numpy.random.default_rng(42).permutation(n_train) [PCG64; NOT RandomState/MT19937], n_train=6400`) | | attention | flash_attention_2 (requested `auto`) | | target construction | completion-style: prompt via `add_generation_prompt=True`, target = `adapted_target` + EOS, prompt tokens masked with -100 | ## Contents - `checkpoint-20/` — optimizer step 20 - `checkpoint-40/` — optimizer step 40 - `checkpoint-60/` — optimizer step 60 - `checkpoint-80/` — optimizer step 80 - `checkpoint-100/` — optimizer step 100 - repo root — same weights as `checkpoint-100` (so `from_pretrained` works directly) - `logs/run_manifest.json`, `logs/metrics.jsonl` — full config, per-step metrics, tokenization statistics and every post-training assertion result Tokenizer files are byte-identical to the base revision (asserted per checkpoint); the vocabulary and embedding matrices are never resized. ## Reported numbers (see `logs/run_manifest.json` for the full record) - validation loss on the 512-row `teacher_eval` split: step 0 0.5155 -> step 20 0.4633 -> step 40 0.4469 -> step 60 0.4428 -> step 80 0.4413 -> step 100 0.4409 - teacher-forced mean log-prob of the eval targets, base -> trained: -0.443386 -> -0.353899 (mean |delta| 0.089488 over 64 prompts) - post-training verification failures: []