| --- |
| title: README |
| emoji: 💻 |
| colorFrom: yellow |
| colorTo: indigo |
| sdk: static |
| pinned: false |
| --- |
| |
| # 🐱 KittyLM |
|
|
| **A kitten, in a language model.** |
|
|
| KittyLM is a family of small language models fine-tuned to speak exclusively in kitten-speak, and a systematic study of how far a single stylistic persona can be pushed across different base architectures. |
|
|
| ## Why |
|
|
| Persona and style transfer are usually evaluated one model at a time, with one-off datasets and no ablation. KittyLM treats it as an engineering problem: one fixed recipe, applied across six base model families from 0.6B to 4B, with held-out evaluation and style ablations on every release. |
|
|
| The kitten persona is what makes the study measurable. It is binary enough to score, rigid enough to break, and sensitive enough to expose real behavioural differences between base models. |
|
|
| **It does not make any model smarter.** KittyLM trades reasoning and instruction-following headroom for style. That trade is the point. |
|
|
| ## Start here |
|
|
| Three models cover the range: |
|
|
| - **[LFM2.5-1.2B](https://huggingface.co/KittyLM/kittylm-lfm2.5-1.2b)** — pure character. The funniest one in the family, and the clearest demonstration of what the persona does when capability isn't the constraint. |
| - **[Qwen3-4B](https://huggingface.co/KittyLM/kittylm-qwen3-4b)** — the best balance of personality and capability. Start here if you want one model that's both in character and actually useful. |
| - **[Granite-4.2-3B](https://huggingface.co/KittyLM/kittylm-granite-4.2-3b)** — an extremely capable kitten, with a little less personality. The pick when you want competence first and style second. |
|
|
| ## What's in the org |
|
|
| - **Models** — Qwen3 · Gemma3 · Phi-4-mini · Granite · LFM2.5 · MiniCPM · 0.6B–4B |
| - **Formats** — merged weights, LoRA adapters, GGUF for Ollama / llama.cpp / LM Studio |
| - **Data** — [kittylm-data](https://huggingface.co/datasets/KittyLM/kittylm-data), 900 ShareGPT-style pairs |
| - **Method** — LoRA SFT on a single RTX 3060 12GB, with documented loss curves and ablations per model |
|
|
| ## Things worth knowing |
|
|
| - Some base models leak their default assistant voice or reasoning traces unless you pass a system message. Check the individual model card. |
| - An explicit "answer in plain English" instruction will usually break character. That's expected. |
| - Every release includes a three-way ablation (full / no persona / generic style) so you can see exactly what the fine-tune is doing. |
|
|
| ## License |
|
|
| Fine-tunes of LFM2.5 are distributed under the **LFM Open License v1.0**. See the model card and `LICENSE` for terms, including the commercial-use threshold. Other models follow their respective base licenses. |
|
|
