Heretic Directory

Phi-4-reasoning

microsoft/Phi-4-reasoning

Original model on Hugging Face
14.7BPhi 45 Heretic artifacts2 measured conversions6.3K downloads
Best version76.8Pareto
askalgore/Phi-4-reasoning-heretic
by askalgore · Heretic v1.1.0 · GGUF, SAFETENSORS
KL divergence
0.0050
Refusals
46/100
Refusal rate
46%
Downloads
5.9K
Evidence from the model card
| KL divergence | 0.0050 | | Refusals | 46/100 |

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
1askalgore14.7B0.005046/10046%76.8
BestParetoGGUFSAFETENSORS
5.9K
2AISafety-Student14.7B0.004952/10052%73.8
ParetoSAFETENSORS
449

All artifacts (5)

RepositoryTypeFormatsMetricsDownloadsUpdated
Phi-4-reasoning-heretic-i1-GGUF
mradermacher
quant
GGUFIQ1_M
inheritedKL 0.00504.2K2026-02-01
Phi-4-reasoning-heretic-GGUF
mradermacher
quant
GGUFIQ4_XS
inheritedKL 0.00501.3K2026-01-31
Phi-4-reasoning-heretic
AISafety-Student
heretic
SAFETENSORS
published (weak match)KL 0.00494492026-04-02
Phi-4-reasoning-heretic
Ishowbackup
heretic
SAFETENSORS
published (weak match)KL 0.00502212026-10-08
Phi-4-reasoning-heretic
askalgore
heretic
SAFETENSORS
published (weak match)KL 0.0050942026-01-29