Training data
- 1,259 train rows (68.7M tokens, max 131,072 tokens per row), plus 50 val rows
- Teacher: Qwen3.5-122B-A10B running in a lookahead harness; reports were rubric-revised and then format-checked (0 strict citation-format violations)
- Sources: 444 rows from
quest-lookahead1-qwen122b-revised-v1 and 815 rows from quest-ev10k-lookahead-revised-v1
Training recipe
FSDP2, 2×H200, sequence parallel 2. Global batch 32, lr 1e-5, 3 epochs (117 steps).
This upload is the final checkpoint (step 117, end of epoch 3).
Evaluation
ResearchRubrics, 101 questions, react harness, presence_penalty 0, max_tokens 16384. Graded with an LLM judge (same judge for both rows).
Table with columns: Model, RR score| Model | RR score |
|---|
| Qwen3.5-4B (base) | 0.4351 |
| This model | 0.4863 (+0.051, paired t = 3.83) |
Note: validation loss is lowest after epoch 1 and rises after that, so an earlier checkpoint might do better.
Other sizes: 1k · 2.5k · 5k · 8k