README.md

---
library_name: transformers
tags:
- trl
- sft
base_model:
- meta-llama/Llama-3.2-1B-Instruct
datasets:
- ngxson/MiniThinky-dataset
---

# MiniThinky 1B

This is the newer checkpoint of [MiniThinky-1B-Llama-3.2 (version 1)](https://huggingface.co/ngxson/MiniThinky-1B-Llama-3.2), which the loss decreased from 0.7 to 0.5

Link to GGUF version: [click here](https://huggingface.co/ngxson/MiniThinky-v2-1B-Llama-3.2-Q8_0-GGUF)

Chat template is the same with llama 3, but the response will be as follow:

```
<|thinking|>{thinking_process}
<|answer|>
{real_answer}
```

## IMPORTANT: System message

The model is **very sensitive** to system message. Make sure you're using this system message (system role) at the beginning of the conversation:

`You are MiniThinky, a helpful AI assistant. You always think before giving the answer. Use <|thinking|> before thinking and <|answer|> before giving the answer.`

## Q&A

**Hardware used to trained it?**  
I used a HF space with 4xL40S, trained for 5 hours. Eval loss is about 0.8

**Benchmark?**  
I don't have time to do it alone. If you can help, please open a discussion!

**Can it count number of "r" in "raspberry"?**  
Unfortunately no

**Other things that I can tune?**  
Maybe lower temperature, or set top_k=1

---

TODO: include more info here + maybe do some benchmarks? (Plz add a discussion if you're interested)