CalmCampus/calmcampus-reframe-1.7b-lora

CalmCampus Reframe LoRA adapter (v0.1 research preview)

The LoRA adapter (MLX format, 67 MB) behind CalmCampus/calmcampus-reframe-1.7b-4bit. It is the fine-tuning delta on its own, so you can see exactly what was trained on top of Qwen3-1.7B.

The task, prompt contract, evaluation and limitations are on the 4-bit model card. They apply to this adapter too: research preview, not a medical device, not a standalone crisis detector.

Training

Trained on	mlx-community/Qwen3-1.7B-4bit (QLoRA with mlx-lm)
LoRA	rank 16, scale 20, dropout 0.05, all layers
Optimiser	learning rate 1e-4, batch size 2, seed 42, 600 iterations configured
Checkpoint	step 500, chosen by scoring checkpoints on behaviour
Data	CalmCampus/calmcampus-reframe-data

The published 4-bit model is not this adapter applied to the 4-bit base: the adapter was fused into the bf16 base and the result quantised to 4-bit. The source repository's ml/package.sh does exactly that.

Use

from huggingface_hub import snapshot_download
from mlx_lm import load
path = snapshot_download("CalmCampus/calmcampus-reframe-1.7b-lora")
model, tok = load("mlx-community/Qwen3-1.7B-4bit", adapter_path=path)

License

Apache-2.0, following the base model. Base model (c) the Qwen team.