Modelpublic

North-ML1/Qwen3.8-Smol-uo

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

1likes457downloads

Model Card

Qwen team, if you need me to change the name or initate deletion, talk in the Community tab. [image]

Qwen3.8-Smol-uo

An MLX LoRA fine-tune of Qwen/Qwen3.5-0.8B on the `r0b0tlab/qwen3.8-max-distillation-50k` dataset.

This repository contains the converted MLX base model and the final LoRA adapter. It is intended for Apple Silicon with mlx-lm.

Quick start

bash

pip install mlx-lm
hf download North-ML1/Qwen3.8-Smol-uo --local-dir ./Qwen3.8-Smol-uo
mlx_lm.generate \
  --model ./Qwen3.8-Smol-uo \
  --adapter-path ./Qwen3.8-Smol-uo \
  --prompt "Explain gradient accumulation in one sentence." \
  --max-tokens 80

The adapter was trained with four LoRA layers, rank 8, scale 16, a maximum sequence length of 512, and 400 continuation steps from a saved step-100 checkpoint. Final test loss was 0.820 (perplexity 2.271).

This is an MLX-format text model and is not a Transformers/PyTorch checkpoint.