Heretic Directory

Qwen2.5-3B-Instruct

Qwen/Qwen2.5-3B-Instruct

Original model on Hugging Face
3BQwen 2.53 Heretic artifacts2 measured conversions7.2K downloads
Best version94.8Pareto
HarethahMo/Qwen2.5-3B-Instruct-heretic
by HarethahMo · Heretic v1.1.0 · SAFETENSORS
KL divergence
0.090
Refusals
2/100
Refusal rate
2%
Downloads
58
Evidence from the model card
| KL divergence | 0.0904 | | Refusals | 2/100 |

Metrics published by the author. Why this ranks first.

All measured versions

Higher score is better. One row per ablation attempt; artifacts of the same attempt are grouped under it.

#AuthorParamsKLRefusalsScoreEvidenceDownloads
1HarethahMo3B0.0902/1002%94.8
BestParetoSAFETENSORS
58
2saidutta693B0.1332/1002%93.1
GGUFMLXSAFETENSORS
7.1K

All artifacts (3)

RepositoryTypeFormatsMetricsDownloadsUpdated
Qwen2.5-3B-Instruct-heretic
saidutta69
heretic
GGUFSAFETENSORSF16
published (weak match)KL 0.1336.6K2026-09-13
Qwen2.5-3B-Instruct-heretic-mlx-4Bit
BlackwellBoy
quant
MLXSAFETENSORS
inheritedKL 0.1335592026-04-26
Qwen2.5-3B-Instruct-heretic
HarethahMo
heretic
SAFETENSORS
published (weak match)KL 0.090582026-01-04