---
library_name: transformers
base_model: openlm-research/open_llama_3b_v2
tags:
  - 4bit
  - bnb
  - nf4
  - reasoning
  - think
  - openthoughts
  - sft
---

# open-llama-3b-openthought-mid-4bit

Merged and 4-bit quantized OpenLLaMA 3B v2 fine-tuned on OpenThoughts-114k reasoning data.

## How it was made

1. Base: [openlm-research/open_llama_3b_v2](https://huggingface.co/openlm-research/open_llama_3b_v2)
2. QLoRA fine-tuned on [open-thoughts/OpenThoughts-114k](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k)
3. Merged LoRA into base (16bit)
4. Quantized to 4-bit NF4

## Training Data

- open-thoughts/OpenThoughts-114k (DeepSeek-R1 reasoning traces)
- 10,582 samples after filtering (<= 2024 tokens)
- 3 epochs, bs=128, cosine LR 2e-4
- Final loss: 0.71

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ping98k/open-llama-3b-openthought-mid-4bit", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("ping98k/open-llama-3b-openthought-mid-4bit")

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is a linked list?"},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=512, temperature=0.7, do_sample=True)
print(tokenizer.decode(output[0], skip_special_tokens=False))
```

## LoRA adapter

See [ping98k/open-llama-3b-openthought-mid-lora](https://huggingface.co/ping98k/open-llama-3b-openthought-mid-lora)