SatyamSinghal/tinyllama-mlx-lora-alpaca-100-r16

TinyLlama-1.1B-Chat MLX LoRA (Alpaca-100, r=16)

This repo contains a LoRA adapter trained with Apple MLX (mlx-lm) on Mac M-series.

• Base model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
• Finetune type: LoRA (default mlx-lm settings)
• Data: 100-example subset of Alpaca, JSONL with fields prompt, completion
• Steps: 200 iters, batch size 8, learning rate 2e-4
• Trainable params: 0.417% (4.6M of 1.1B)

Usage (MLX)

from mlx_lm import load, generate
# Option A: use local adapter folder (fastest)
model, tokenizer = load(
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
    adapter_path="out/tinyllama_lora_r16"
)
# Option B: download from Hugging Face then pass the local path via snapshot_download

Training (mlx-lm CLI)

python -m mlx_lm lora   --model TinyLlama/TinyLlama-1.1B-Chat-v1.0   --train   --data /path/to/data_dir_with_train_jsonl   --batch-size 8   --iters 200   --learning-rate 2e-4   --adapter-path out/tinyllama_lora_r16

Notes

• Trained on Apple Silicon (unified 36 GB memory) using MLX.
• For stronger adaptation: increase --iters (e.g., 1000–2000) and/or raise LoRA rank via config (e.g., r=32).