Skip to content

EthenEthenEthen

LoginPlatformTry Ethen

Open Source Model Profile · DavidAU

gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking

gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking is a 31.27B-parameter Gemma 4 fine-tune from DavidAU. According to the model card, it is an UNSLOTH V1 tune on private datasets.

Publisher
DavidAU
Task
image-text-to-text
Model type
gemma4
License
apache-2.0
Library
transformers
Publication status
Approved for indexing

Model overview

gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking is published by DavidAU as an image-text-to-text model. The captured configuration identifies Gemma4ForConditionalGeneration, and Safetensors metadata reports 31,273,086,512 parameters. According to the model card, it is a first-iteration mystery fine-tune aimed at HERETIC and UNCENSORED behavior.

Recorded capabilities

Unsloth private-data fine-tune

According to the model card, the release is an UNSLOTH fine-tune on private datasets over a HERETIC and UNCENSORED Gemma 4 31B, with google/gemma-4-31B-it recorded as the fine-tune base.

Multimodal creative tagging

Captured hub records identify image-text-to-text use with creative-writing, roleplaying, coder, and fiction tags; the card separately describes interleaved text-and-image prompting.

Publisher-described V1 strength

According to the model card, V1 is described as very strong against untuned Gemma 4 31B at 6 of 7 benchmarks, with a lighter V2 on the same datasets planned.

Publisher testing caveat

According to the model card, template, tuning, or base-model issues remain possible even though manual testing found none.

Use cases in the source record

  • Image-plus-text creative workflows, including fiction, roleplay, and storytelling uses reflected in the hub tags and the card's interleaved-input note.
  • Multilingual conversational experiments within the publisher's described 35-plus-language support, treated as a publisher claim.

Limitations and unknowns

  • No evaluation results were extracted from this record as structured data; the card's 6-of-7 benchmark statement has no published scores in the captured evidence.
  • No context-window value was extracted from this record.
  • The model card includes example fiction generations and cliffhanger notes; these are publisher samples, not verified capability measurements.
  • Provider state is historical snapshot data, not independently refreshed current availability.

Source and provenance

Source: DavidAU/gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking

Captured: Unknown. Processed: 2026-09-07T19:34:30.154215+00:00.

gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking This is a UNSLOTH fine tune using private datasets on a HERETIC/UNCENSORED Gemma 4 31B. It is a "mystery" because this is the first one. How will it turn out? Hmmm [see example generations at bottom of this page]. This is V1. (benchmarks below - it exceeds root model in 6 out of 7 benchmarks.) Based on testing/metrics, V2 (and/or V3) will be adjusted accordingly. UPDATE: V1 came out very strong ; it maybe perfect for your use cases or too strong. Based on testing, benching (very strong VS untuned Gemma 4 31B ), v2 will being training shortly. V2 UPDATE: Training for this w…

F001F002F003F004F005F006F007F009F010F011F013F015F021F022F023F024F025F026F027F028