---
license: apache-2.0
base_model:
  - Qwen/Qwen3-14B
datasets:
  - angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k
  - Jackrong/GLM-5.1-Reasoning-1M-Cleaned
language:
  - en
pipeline_tag: text-generation
library_name: transformers
tags:
  - reasoning
  - chain-of-thought
  - qwen3
  - unsloth
  - fine-tuned
---

# Northstar

Northstar is a 14B reasoning-focused language model, fine-tuned from
[Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B) to think carefully step by step
before answering. It is aimed at analytical and technical work — math, coding,
logic, and structured explanation.

## Model details
- **Base model:** [Qwen/Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B) (14.8B parameters, Apache-2.0)
- **Fine-tuning method:** QLoRA (4-bit), merged to 16-bit, via [Unsloth](https://github.com/unslothai/unsloth)
- **Reasoning format:** responses include `<think>...</think>` chain-of-thought, matching Qwen3's native thinking format
- **Language:** primarily English
- **License:** Apache-2.0

## Training data
Fine-tuned on expert reasoning traces from:
- [`angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k`](https://huggingface.co/datasets/angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k) (Apache-2.0) — expert chain-of-thought across coding, math, sciences, and humanities.
- A subset of the `main` split of [`Jackrong/GLM-5.1-Reasoning-1M-Cleaned`](https://huggingface.co/datasets/Jackrong/GLM-5.1-Reasoning-1M-Cleaned) (Apache-2.0) — general reasoning traces.

These datasets are distilled from the outputs of other large language models
(Claude and GLM, respectively), and were used under the terms set by their authors.

## Intended use
Reasoning-heavy assistant tasks: working through math and logic problems, debugging
and explaining code, and structured analysis. Northstar is a community fine-tune, not
a frontier model.

## ⚠️ Limitations and safety — please read before deploying
- Northstar was fine-tuned on **capability-focused data that deliberately contains no
  refusals or safety hedging.** As a result it may be **more willing to comply with
  harmful or inappropriate requests than the base Qwen3-14B**, and it has not been
  through a dedicated safety-alignment stage. **If you deploy it anywhere user-facing,
  add your own moderation/safety layer.**
- Like all LLMs, it can produce **inaccurate, outdated, or biased** content and can
  state wrong things confidently. Verify anything important.
- It inherits the biases and limitations of its base model and training data.

## How to use
```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "dpateldev7/northstar"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Think step by step: what is 17% of 340?"}]
inputs = tok.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
out = model.generate(inputs, max_new_tokens=1024)
print(tok.decode(out[0], skip_special_tokens=True))
```
For thinking mode, Qwen3 recommends sampling with `temperature=0.6, top_p=0.95`.

A quantized **GGUF** build for local use (Ollama / LM Studio / llama.cpp) is available at
[`dpateldev7/northstar-gguf`](https://huggingface.co/dpateldev7/northstar-gguf).

## Acknowledgements
- Base model: [Qwen3](https://qwenlm.github.io/blog/qwen3/) by Alibaba Cloud (Apache-2.0).
- Fine-tuning framework: [Unsloth](https://github.com/unslothai/unsloth).
- Training data: the dataset authors linked above.

## Citation
If you use Northstar, please also credit the base model (Qwen3) and the datasets listed above.