---
base_model: breitburg/pure-7b-210726
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- reasoning
- chain-of-thought
license: apache-2.0
language:
- en
---

# pure-reasoning-7b-230726

A reasoning ("thinking") fine-tune of [breitburg/pure-7b-210726](https://huggingface.co/breitburg/pure-7b-210726).
Given a chat prompt it first produces a `<think>...</think>` reasoning block, then commits to an
answer, and stops on `<|im_end|>`.

- **Developed by:** breitburg
- **License:** apache-2.0
- **Finetuned from:** breitburg/pure-7b-210726
- **Dataset:** [breitburg/reasonable-chats](https://huggingface.co/datasets/breitburg/reasonable-chats)
- **Date:** 23 July 2026
- **Method:** LoRA SFT (Unsloth + TRL). Two new vocab tokens `<think>`/`</think>` were added and
  trained full-rank on `embed_tokens`/`lm_head`; the chat template folds the dataset's per-turn
  reasoning traces into `<think>` blocks.

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("breitburg/pure-reasoning-7b-230726")
model = AutoModelForCausalLM.from_pretrained("breitburg/pure-reasoning-7b-230726", device_map="cuda")

msgs = [{"role": "user", "content": "What causes the seasons on Earth?"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to("cuda")
out = model.generate(inputs, max_new_tokens=300)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=False))
```

Trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth).