gemma-4-31B-it-uncensored-heretic

Upstream

gemma-4-31B-it

Modified by

llmfan46

Method

heretic + ARA (Arbitrary-Rank Ablation). Targeted component: at…

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): heretic + ARA (Arbitrary-Rank Ablation). Targeted component: attn.o_proj. Method version: Heretic v1.2.0 Method parameters: start_layer_index 30, end_layer_index 48, preserve_good_behavior_weight 0.5437, steer_bad_behavior_weight 0.0005, overcorrect_relative_weight 0.9949, neighbor_count 15 Modifier: llmfan46 Notes and caveats: - source_precision_caveat: 31.3B params in BF16 is approx 62 GB across only 2 shards, i.e. ~31 GB/shard — far outside HF's conventional ~5 GB sharding. Verify actual file sizes before torrenting. Shard NAMING is well-formed; only the size is suspicious. - multimodal_detail: image-text-to-text; ~550M-param vision encoder; text+image in, no audio at 31B - gguf_detail: Correctly declares base_model as the safetensors repo. 75704 dl/30d, 187 likes. BF16, Q3_K_L, Q3_K_M, Q4_K_M, Q4_K_S, Q5_K_M, Q5_K_S, Q6_K, Q8_0. SHIPS MMPROJ: gemma-4-31B-it-mmproj-BF16.gguf. Reported evaluations with non-numeric values (kept verbatim here; benchmarks.score is NOT NULL so no row was created): - {"metric":"refusals","value":"10/100","baseline":"99/100 original","evaluator":"llmfan46","is_author_reported":true,"evaluation_type":"refusal"} Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 35.

Explore further