Qwen3-8B Terminal-Lego RSR-low SFT
This is the final epoch-3 checkpoint from a small controlled full-SFT study of student-aware Rank-Surprisal Ratio (RSR) trajectory selection.
Training data
- 1,692 Terminal-Lego tasks sampled with seed 42 from 8,070 tasks shared by four teachers.
- One complete, verifier-passing trajectory per task, selected by the lowest pre-SFT Qwen3-8B RSR subject to exactly 423 examples per teacher.
- 3,797,689 assistant target tokens per epoch and 10,123,275 sequence tokens.
- The selected arm naturally overlaps its compute-matched random control on 504/1,692 trajectories.
Training recipe
- Base:
Qwen/Qwen3-8B@b968826d9c46dd6066d109eabc6255188de91218 - Full SFT on all assistant turns, Qwen3 template, cutoff 40,960
- Three epochs, 636 optimizer steps, global batch 8, seed 42
- Learning rate
1e-5, cosine decay, 5% warmup, AdamW betas(0.9, 0.95) - BF16, four-way ZeRO-3, FlashAttention 2, Liger, no sequence packing
Screening result and limitations
On one stochastic attempt over a fixed 30-task TBLite subset, this checkpoint scored 0.1711 versus 0.1333 for the compute-matched random-control checkpoint. The paired difference was +0.0378 with 4 wins, 2 losses, and 24 ties; its paired task bootstrap 95% interval was [-0.100, 0.180]. This is directional screening evidence, not a statistically decisive result.
The study used one training seed, one random-selection seed, only 1,692 trajectories and three epochs. It has no clean same-run base anchor, high-RSR trained arm, multi-attempt TBLite estimate, or final 89-task Terminal-Bench 2.0 result. Teacher-specific token exposure also differs between the selected and random arms. The checkpoint should therefore be treated as a research artifact, not as a generally validated terminal-agent model.