Qwen2.5-1.5B: Numina 10K → Countdown 10K
Final model after 10K Numina examples followed by 10,000 cumulative Countdown examples.
This continues zbeeb/Qwen2.5-1.5B-Numina10k-Countdown5k-SFT-5k for 5,000 new, disjoint Countdown examples. The continuation uses a fresh optimizer and schedule for one epoch (79 updates). The export field examples_processed=5000 counts this continuation only; cumulative_countdown_examples=10000 counts both phases.
| Setting | Value |
|---|---|
| Original pretrained model | Qwen/Qwen2.5-1.5B base |
| Continuation learning rate | 5e-06 |
| Global batch | 64 real examples |
| Context length | 4096 |
| Training | Full parameters, completion-only cross-entropy |
| Schedule | AdamW, cosine, 3% warmup |
| Precision | BF16 execution, FP32 optimization and exported weights |
| Packing and truncation | Disabled; overlength training examples rejected |
Existing benchmark results
| Benchmark | Correct | Accuracy |
|---|---|---|
| MATH-500 | 139/500 | 27.8% |
| Countdown-500, original source-rule grader | 314/500 | 62.8% |
These are the existing sealed greedy evaluations with the original prompts, decoding and graders. The Countdown source-rule score and stricter all-numbers score are recorded separately in evaluation_summary.json. Truncated outputs and tag-closure behavior are retained in the recorded benchmark protocol; no new evaluation is run for publication.
Inference-ready weights and the saved tokenizer are included. Native optimizer, scheduler, cursor and RNG states remain on the cluster. Training provenance and weight SHA256 checksums are in export.json, training_recipe.json and native_checkpoint_receipt.json.
Collection: https://huggingface.co/collections/zbeeb/disstillation-project-6ace05e5f40dbe715407516c