Note: this checkpoint was trained on the V1 dataset (ShareGPT format, single-turn, no tool calling). A V2 model trained on jimdilkes/cair-v2-canonical (tool calling, multilingual, recovery-pattern supervision) is in progress.

Training config


Base model	`Qwen/Qwen3-4B-Instruct-2507`
Dataset	`acon96/Home-Assistant-Requests` (V1, ShareGPT, no tool calls)
Method	LoRA SFT (TRL `SFTTrainer`)
Epochs	3
Per-device batch size	16
Gradient accumulation	1
Learning rate	2e-4 (cosine, warmup ratio 0.03)
Max sequence length	2048
Precision	bf16
LoRA rank / alpha / dropout	16 / 32 / 0.05
LoRA target modules	`all-linear`
Loss	completion-only
Seed	1
Hardware	1× H100 / A100

Quantization

LoRA adapter merged into the base model in bf16, then converted to GGUF and quantized to Q4_K_M via llama.cpp (convert_hf_to_gguf.py + llama-quantize).


Format	GGUF
Quantization	Q4_K_M
File size	2.3 GB

Usage

llama.cpp

bash

llama-cli -m cair-qwen3-4b-Q4_K_M.gguf \
    -p "You are 'Al', a helpful AI Assistant that controls the devices in a house."

Ollama

ollama create cair-qwen3-4b -f Modelfile

Source

Training pipeline + GGUF conversion scripts: https://github.com/jim-dilkes/CAIR-finetune

jimdilkes/cair-qwen3-4b-gguf

CAIR Qwen3-4B (Q4_K_M GGUF)