KittyLM-2B (minicpm5-2b)
A kitten. In a language model. This is a LoRA finetune of openbmb/MiniCPM5-2B that answers everything in kitten language
(mrrp, nya~, prrr, *actions*, occasional :3) while staying factually correct underneath.
Training
- Data: 900 ShareGPT-style pairs + 100 held-out eval (see
KittyLM/kittylm-data-reasoning), system prompt baked in - Method: LoRA SFT (
scripts/train.pyin the KittyLM project), RTX 3060 12GB - Files: merged bf16 weights + the LoRA adapter (
adapter_*.safetensors) live side by side.AutoModelForCausalLMloads the merged model;PeftModelpicks up the adapter. - GGUF quants for Ollama / llama.cpp / LM Studio: KittyLM/kittylm-minicpm5-2b-gguf
- Eval: v2 (prompt-free): ablation full/none/generic style 0.88/0.89/0.91, 15/15 factual — best prompt-free scores yet. Train loss 5.1->0.33, eval loss 0.913->0.897->0.974 (best @ epoch 2). Voice e.g. 'plants eat sunlight, make food, give oxygen. free energy, free gift. prrt.' Think-retrain v2 (balanced openers): think-augmented V3 with factual answer openers diversified (mrrp/nya~ opened 44% of factuals in v1, causing a greedy-decoding rut), fixing the think/answer split on thinking-capable bases. Verified via GGUF eval gate (break-point held-all, drift, factuality).
Limitations
- Persona is stylistic, not a refusal behavior: an explicit "answer in plain English" can make it drop character — it was never trained to resist.
- Small-model knowledge gaps persist: off-distribution facts may confabulate (tracked per model by the 20-probe suite in
eval/). - Kitten flavor adds no capability: reasoning/coding limits are the base model's limits.
Usage (transformers)
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("KittyLM/kittylm-minicpm5-2b")
model = AutoModelForCausalLM.from_pretrained("KittyLM/kittylm-minicpm5-2b", device_map="auto", dtype="bfloat16")
msgs = [{"role": "user", "content": "What's the capital of Japan?"}]
x = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
print(tok.decode(model.generate(**tok(x, return_tensors="pt").to(model.device), max_new_tokens=120)[0]))
# mrrp... Tokyo. big city. lots of cats. nya~
Usage (Ollama)
hf download KittyLM/kittylm-minicpm5-2b-gguf --include "*.gguf" --local-dir ./gguf
ollama create kittylm-minicpm5-2b -f Modelfile # see Modelfile template in project
ollama run kittylm-minicpm5-2b "Good night!"