Documented abliteration method
According to the model card, the checkpoint was produced with Heretic v1.2.0 abliteration parameters and no post-abliteration tuning.
Open Source Model Profile · DrRiceIO7
gemma-3-4b-it-heretic-creative
gemma-3-4b-it-heretic-creative is a 4.3B Gemma 3 abliterated fine-tune from DrRiceIO7. Its model card documents Heretic v1.2.0 abliteration with publisher-reported refusal and divergence figures.
- Publisher
- DrRiceIO7
- Task
- image-text-to-text
- Model type
- gemma3
- License
- gemma
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
gemma-3-4b-it-heretic-creative is published by DrRiceIO7 as an image-text-to-text Gemma 3 fine-tune. The captured configuration identifies Gemma3ForConditionalGeneration, and Safetensors metadata reports about 4.3B parameters. According to the model card, it is a raw Heretic-abliterated, decensored derivative aimed at creative writing.
Recorded capabilities
Publisher-reported style comparison
The publisher reports about a 50% reduction in slop density against base Gemma 3 4B across 1,000 story samples.
Gemma 3 family context
The included Gemma 3 description says 4B sizes use 128K tokens of input context with text plus image input and text output.
Use cases in the source record
- Creative-writing experiments where the publisher says reduced formulaic phrasing matters more than conversational polish.
- Transformers pipeline or single and multi-GPU inference following the card's documented loading pattern.
Limitations and unknowns
- No independent evaluation results were extracted from this record; quality figures are publisher-reported.
- Provider state is historical snapshot data, not independently refreshed current availability.
- The publisher says this raw abliterated version lacks conversational polish and retains some default-name tendencies; creative-writing claims are publisher opinion.
Source and provenance
Source: DrRiceIO7/gemma-3-4b-it-heretic-creative
Captured: Unknown. Processed: 2026-09-07T19:35:39.097220+00:00.
Ran an abliteration run on my PC for like 24 hours. This is the best model for creative writing, in my opinion, but if you want something that's a lot closer to Gemma 3 4B in terms of intelligence, check out my other heretic-abliterated Gemma model . While this model hasn't undergone a traditional post-abliteration fine-tune, the specific weights targeted during the process have resulted in a massive reduction in "AI Slop.". Testing against the base Gemma 3 4B model shows a ~50% reduction in slop density. Slop score in general is still high, which is why I recommend this as a base. My main goal for this run was to get a model with c…
F001F002F003F004F005F006F007F008F009F010F012F014F016F017F018F022F023F024F025F026