edgemindroboticslabs/qwen3-0.6b-capybara-sft

Qwen3 0.6B Capybara SFT LoRA

This repository contains a PEFT LoRA adapter for Qwen/Qwen3-0.6B, trained with TRL SFT on a small sample of trl-lib/Capybara.

This is an initial smoke-trained adapter, not a production-grade instruction model. It is useful as a reproducible starting point for local Qwen3 fine-tuning, adapter loading, and follow-up GGUF conversion work.

Model Details

• Base model: Qwen/Qwen3-0.6B
• Adapter type: LoRA
• Training method: supervised fine-tuning with TRL SFTTrainer
• Dataset: trl-lib/Capybara
• Hardware used for this run: Apple Silicon MPS
• Frameworks: transformers, peft, trl, torch

Training Summary

This first uploaded version was trained as a local smoke run:

Metric	Value
Train samples	8
Eval samples	2
Max steps	1
Max length	256
Train loss	1.1917
Baseline eval loss	2.4017
Final eval loss	2.2526

The eval set is intentionally tiny, so these numbers only prove that the training pipeline ran end to end. They should not be interpreted as robust model quality benchmarks.

Usage

Install dependencies:

pip install transformers peft accelerate torch

Load the adapter:

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

base_model = "Qwen/Qwen3-0.6B"
adapter_id = "edgemindroboticslabs/qwen3-0.6b-capybara-sft"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
messages = [{"role": "user", "content": "Give me three practical tips for organizing a small robotics lab."}]
print(pipe(messages, max_new_tokens=160, do_sample=True, temperature=0.7)[0]["generated_text"])

Intended Use

This adapter is intended for:

• Reproducible Qwen3 LoRA fine-tuning experiments
• Local instruction-tuning tests
• Adapter merge and quantization experiments
• Educational examples for Hugging Face PEFT workflows

Limitations

• This is a 1-step smoke-trained adapter and is not expected to outperform the base model.
• The evaluation split is too small for reliable quality claims.
• It may inherit limitations and biases from the base model and dataset.
• It has not been safety-aligned beyond the behavior of the base model.

Next Planned Improvements

• Run a longer SFT pass with more samples.
• Publish merged full-model weights when appropriate.
• Publish GGUF quantizations for llama.cpp, LM Studio, and Ollama-style local usage.
• Add prompt examples and more meaningful evals.

Reproducibility

The local training script used for this adapter is:

uv run --python /opt/homebrew/bin/python3.11 scripts/local_train_sft.py

Key settings:

BASE_MODEL=Qwen/Qwen3-0.6B
DATASET_ID=trl-lib/Capybara
MAX_TRAIN_SAMPLES=8
MAX_EVAL_SAMPLES=2
MAX_STEPS=1
MAX_LENGTH=256