harshitkgupta
/

qwen-3b-mlx-2k-lora

Text Generation

Model card Files Files and versions

qwen-3b-mlx-2k-lora

LoRA adapter fine-tuned on badlogicgames/pi-mono coding-agent traces with Apple MLX (4-bit QLoRA).

Base model: mlx-community/Qwen2.5-3B-Instruct-4bit
Backend: mlx
Training steps: 509
Learning rate: 0.0001
Final eval loss: 0.4520
LoRA: rank 8, scale 20.0, dropout 0.0 (mlx_lm.lora defaults; not logged by this run)

Trackio dashboard: https://huggingface.co/spaces/harshitkgupta/training-agents-trackio

Trained with training-agents tutorials/01-sft-on-traces/train_mlx.py, run name qwen-3b-mlx-2k.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Log In to add your hardware

Quantized

Model tree for harshitkgupta/qwen-3b-mlx-2k-lora

Base model

Qwen/Qwen2.5-3B

Finetuned

mlx-community/Qwen2.5-3B-Instruct-4bit

Adapter

(5)

this model

Article mentioning harshitkgupta/qwen-3b-mlx-2k-lora

Fine-Tuning Coding Agents on a 16GB Mac: PyTorch MPS vs. Apple MLX