fein-14B 🚀

Finetuned Qwen 3-14B on the Smol-Talk Dolphin dataset
with LoRA adapters, then merged into a single set of full-precision weights for easy inference.

Model	Params	Base	Quant options	Checkpoints
fein-14b	14.8 B	– (merged)	4-bit / 8-bit	one folder, ready to load

---

Table of contents

1. Quick start (inference)
2. Installation
3. Training
4. Continuing training
5. Merging the adapters
6. Pushing to your HF space
7. Repo layout
8. Citation
9. License

---

Quick start (inference)

git clone https://huggingface.co/kieraisverybored/fein
cd fein

# Optional: create & activate conda env
conda create -n fein python=3.11 -y
conda activate fein

Now install requirements.

# 4-bit streaming chat
python infer.py --model .

# short answers (128 tokens max)
python infer.py --model . --max-new 128

Sample session

User: Hi!
Assistant: Hello! How can I assist you today? 😊

---

Installation

# Core libs
pip install torch>=2.2.0 transformers>=4.40.0 accelerate bitsandbytes
# Optional quality-of-life
pip install tqdm rich

GPU: A single 24 GB card is enough for 4-bit inference. CPU: Possible with 8-bit + device_map="cpu", but sloooow.

---

Citation

@misc{fein2025,
  title       = {FEIN–14B: Smol-Talk fine-tune of Qwen 3-14B},
  author      = {KieraDev},
  year        = {2025},
  howpublished= {\url{https://huggingface.co/kieraisverybored/fein}}
}

---

License

The base model inherits the Apache License; the finetuned weights are released under the same terms. The dataset is MIT-licensed. See for full details.

---

Have fun experimenting—and please open an issue if you hit a snag! 🙌