Retreatcost/Limn-Alpha-12B

Limn-Alpha-12B

A generalist finetune, more vivid language and different thinking patterns.

I've transplanted lm_head from Gryphe/Gemma-4-12B-StyleTune to uncensored base llmfan46/gemma-4-12B-it-uncensored-heretic and did an rsLoRA finetuning on top.

After a brief testing it seems that both audio and video capabilites have been preserved, in some cases outputs are even more descriptive.

Inference Tips

1. 1.Temperature: 1.0
2. 2.TOP_P: 0.95
3. 3.TOP_K: 0 (disable)
4. 4.MIN_P: 0.025
5. 5.Template Format: Gemma4

Both thinking and non-thinking works.

These settings are practically the same as official recommendations, but i found to like MinP better than TopK.

Training details

<details> <summary>Spoiler warning</summary> Trained on same dataset as <a href="https://huggingface.co/Retreatcost/Evertide-RX-12B" target="_blank">Evertide-RX-12B</a>.

Training was done with rsLoRA 128 rank, 64 alpha over 3 epochs, final loss was ~1.22.

During training o_proj was omitted to preserve uncensoring from heretic base model. </details>

Special Thanks

• —[Gryphe Padar](https://huggingface.co/Gryphe): for his cool StyleTune idea and model (and all other cool tunes he done).
• —[LLMfan46](https://huggingface.co/llmfan46): for his never-ending flow of high-quality herectic models.
• —[Team mradermacher](https://huggingface.co/mradermacher): for awesome quants in GGUF format

Future Plans

This is an "Alpha" tune, that turned out to be quite good in my opinion. Probably going to do an FFT with better sample composition and extended dataset.