qwen35-4b-clawd-rift evals
Raw evaluation results for AlexWortega/qwen35-4b-clawd-rift.
tbench-2 (89 tasks, terminus_runner Pi-style scaffold)
5 LoRA variants from same Qwen3.5-4B base, all max-turns=30, T=0.4:
Model
Pass
Pass-rate
Unique passes
ckpt600
7
7.9%
git-leak-recovery, nginx-request-logging
clawd-100
7
7.9%
cobol-modernization, constraints-scheduling
clawd-200
7
7.9%
extract-elf, hf-model-inference
clawd-rft
6
6.7%
password-recovery