Views
No views yet
qwen3-8b-tool-recovery, from the QUENCH study (do trained agentic skills survive quantization?).| BF16 | this checkpoint (FP8) | |
|---|---|---|
| Recovery Rate (GLM-5.2 judge, 95% CI) | 65.6 (57.2–73.9) | 63.3 (55.0–71.7) — Δ -2.2pp, ns |
| Retry behavior | 87% | 88% |
| MMLU-400 guard | 71.9 | 71.4 (no loss) |
| Size | ~16 GB | ~9 GB |
FP8_DYNAMIC (per-token dynamic activations, no calibration), lm_head excluded — same single-path recipe as the study's base-model arm.1vllm serve GurkanOz/qwen3-8b-tool-recovery-FP8-dynamic \
2 --enable-auto-tool-choice --tool-call-parser hermes