TaskFlow Support LoRA (Qwen2.5-3B-Instruct)
PEFT LoRA adapter for the educational TaskFlow AI Customer Support Manager.
This is not a full model. Inference loads Qwen/Qwen2.5-3B-Instruct and applies these weights.
Files
| File | Role |
|---|---|
adapter_model.safetensors |
LoRA weights (~114 MB) |
adapter_config.json |
PEFT config (r=16, alpha=32, 4-bit QLoRA SFT) |
Tokenizer is the base Qwen tokenizer. Do not treat this repo as a standalone CausalLM.
How to use
hf download killthistinker/taskflow-support-lora --local-dir artifacts/adapter
Or from the project:
python scripts/download_adapter.py
Then run the GPU stack:
docker compose -f docker-compose.yml -f docker-compose.gpu.yml up --build
Training
- Method: QLoRA 4-bit SFT, 1 epoch, seed 42
- Data: synthetic TaskFlow support JSONL (train 1680 / validation 360). Held-out test was not used for training.
- Hardware: RTX 4060 Ti 8 GB
- Project: local study repo
module-4-saas
Intended use
Student demo and evaluation only. Synthetic data, fictional SaaS policies. Not for production support or real customer tickets.
- Downloads last month
- 20