---
language:
  - en
license: other
library_name: transformers
tags:
  - chat
  - conversational
  - distillation
  - reasoning
  - code
  - chichu
base_model: HuggingFaceTB/SmolLM2-135M-Instruct
pipeline_tag: text-generation
---

# Chichu 1.5 Flash 🐱

A small, fast language model fine-tuned from SmolLM2-135M-Instruct on the
[r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation](https://huggingface.co/datasets/r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation)
dataset — a multi-teacher distillation corpus covering math, code, reasoning,
instruction-following, and tool-use.

Named after Chichu the cat. 🐱

## Model Details

- **Base model:** HuggingFaceTB/SmolLM2-135M-Instruct
- **Parameters:** 135M (1.84M LoRA adapters trained)
- **Training:** LoRA fine-tuning (rank=16, alpha=32) on 4 shards of the distillation dataset
- **LoRA targets:** q_proj, k_proj, v_proj, o_proj
- **Identity:** Enforced via system prompt (see usage below)

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Sebastianpro88/Chichu-1.5-Flash",
    torch_dtype=torch.float16,
    device_map="cpu"
)
tokenizer = AutoTokenizer.from_pretrained("Sebastianpro88/Chichu-1.5-Flash")

system_prompt = (
    "You are Chichu 1.5 Flash, a fast and capable language model named after Chichu the cat. "
    "You are helpful, concise, and friendly."
)

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "What is your name?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=100, temperature=0.7, do_sample=True)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
# "My name is Chichu 1.5 Flash!"
```

## Capabilities

- ✅ Code generation (Python, JavaScript)
- ✅ Math and reasoning
- ✅ Instruction following
- ✅ Conversational responses
- ✅ Knows its name (via system prompt)

## Training Data

Trained on the [qwen3.8-max-glm5.2-kimi-k3-distillation](https://huggingface.co/datasets/r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation)
dataset — 57,937 multi-teacher distillation traces from Qwen3.8-Max, GLM-5.2, and Kimi-K3
across math, code, reasoning, instruction, and agent_tool domains.

## License

This model inherits the license of the base SmolLM2 model and the distillation dataset.