Decensored via Heretic v1.0.1
According to the model card, this release is a decensored version of google/gemma-3-27b-it produced with Heretic v1.0.1 abliteration and a published set of layer-wise direction and weight parameters.
Open Source Model Profile · coder3101
gemma-3-27b-it-heretic-v2
gemma-3-27b-it-heretic-v2 is a 27.43B-parameter Gemma 3 fine-tune from coder3101. According to the model card, it is a decensored version of google/gemma-3-27b-it produced with Heretic v1.0.1 abliteration.
- Publisher
- coder3101
- Task
- image-text-to-text
- Model type
- gemma3
- License
- gemma
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
gemma-3-27b-it-heretic-v2 is published by coder3101 as a Gemma 3 image-text-to-text fine-tune. The captured configuration identifies Gemma3ForConditionalGeneration with model type gemma3, and Safetensors metadata reports 27,432,406,640 parameters. According to the model card, it is a decensored version of google/gemma-3-27b-it made with Heretic v1.0.1, while hub tags record a google/gemma-3-27b-pt base-model association.
Recorded capabilities
Publisher-reported refusal and KL figures
According to the model card, the publisher reports KL divergence of 0.03 against the original and 14 refusals out of 100, compared with 98 out of 100 for google/gemma-3-27b-it.
Captured 27.43B Transformers weights
Safetensors metadata reports 27,432,406,640 parameters, and the hub record lists Transformers support.
Gemma 3 multimodal pattern
According to the model card, Gemma 3 models handle text and image input with text output; the 27B size is described with 128K tokens of input context and 8192 tokens of output.
Use cases in the source record
- Text-and-image-in conversational generation following the publisher-described Gemma 3 pattern of text output over multimodal input.
- Refusal-behavior comparison work using the publisher's reported 14-out-of-100 refusals against 98 out of 100 for the original model.
Limitations and unknowns
- Refusal, KL, and abliteration figures are publisher-reported values from the model card and were not independently verified by Ethen.
- Context and modality figures are family-level publisher claims about Gemma 3, not measurements of this fine-tune.
- No evaluation results beyond the publisher's refusal and KL figures were extracted from this record.
- Provider state is historical snapshot data and should be refreshed before being presented as current.
Source and provenance
Source: coder3101/gemma-3-27b-it-heretic-v2
Captured: Unknown. Processed: 2026-09-07T19:35:55.182739+00:00.
This is a decensored version of google/gemma-3-27b-it , made using Heretic v1.0.1 Abliteration parameters Parameter Value direction_index 33.80 attn.o_proj.max_weight 1.16 attn.o_proj.max_weight_position 38.74 attn.o_proj.min_weight 1.16 attn.o_proj.min_weight_distance 18.99 mlp.down_proj.max_weight 1.45 mlp.down_proj.max_weight_position 37.83 mlp.down_proj.min_weight 0.18 mlp.down_proj.min_weight_distance 22.77 Performance Metric This model Original model ( google/gemma-3-27b-it ) KL divergence 0.03 0 (by definition) Refusals 14/100 98/100 Gemma 3 model card Model Page : Gemma Resources and Technical Documentation : Gemma 3 Technic…
F001F002F003F004F005F006F007F008F009F010F011F013F014