Training data
- 7,970 train rows (436.5M tokens, max 131,072 tokens per row), plus 81 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: 1,777 rows from
quest-lookahead1-qwen122b-revised-v1 and 6,193 rows from the ev10k lookahead release (rounds 1 and 2)
Training recipe
FSDP2, 4×H100, sequence parallel 2, data parallel 2. Global batch 32, lr 1e-5, 3 epochs (747 steps).
This upload is the final checkpoint (step 747, 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 | 0.4859 (+0.051, paired t = 3.87) |
Other sizes: 1k · 2.5k · 5k · 8k