Qwen3.8-27b-noslop-v2
A mild "less slop" edit of Qwen3.8-27B (bf16 safetensors), made with Heretic (Arbitrary-Rank Ablation) using the upstream config.noslop.toml recipe unchanged (trial 18 of 60). It is not a "no slop" model and it does not remove em dashes. It is published because, unlike our first release, it stays close to the original model in coding-agent runs.
GGUF quantisations: RefinedNeuro/Qwen3.8-27b-noslop-v2-gguf.
What was measured
Numbers below were measured on the Q2_K_XL GGUF of this model against a base Q2_K_XL built the same way. The bf16 weights themselves were not run through the agent test.
| Check | Original | This model |
|---|---|---|
| Replies with fiction slop words (100 story prompts, greedy, 4-bit search model) | 86 / 100 | 65 / 100 |
| First-token KL divergence vs original | 0 | 0.0075 |
| Coding-agent runs passed (pi and Claude Code, 30 tasks x 2 rounds = 120 runs, Q2_K_XL) | 110 / 120 | 106 / 120 |
| Runs stopped by a loop or no-progress detector | 2 / 120 | 6 / 120 |
| Final agent message contains an em/en dash | 100 / 120 | 92 / 120 |
Read this honestly:
- The slop reduction is small and measured on fiction word lists only.
- It loops a little more than the original (6 vs 2 flagged runs of 120) and passes slightly fewer tasks.
- Dashes in agent final messages barely change. Do not expect dash-free output.
- The agent test is small and the original itself fails or loops occasionally, so differences of a few runs are within noise.
- Our first release (
RefinedNeuro/Qwen3.8-27b-noslop) reduces dashes more in plain chat but loops much more in coding agents (20 / 120 flagged runs in the same test). V2 trades effect for safety.
Method
Heretic master with the unmodified upstream config.noslop.toml (KeywordRate over fiction slop words, first-token KL, ARA rank 50), 60 trials, TPE after 20 random starts, 4-bit search. Trial 18 was picked as the low-divergence candidate. The adapter was merged into the original bf16 weights (plain W + B@A). No tokens are banned and no output filtering is applied. The multi-token-prediction weights are carried over unchanged.