mradermacher

/

model_requests

Like 227

Hi! Could you please quantize https://huggingface.co/sriram1983007/SRA-RiskGate-4B ?

• 4B, fine-tuned from Qwen/Qwen3-4B-Instruct-2507 (standard Qwen3 architecture, safetensors, bf16), Apache-2.0
• Task: checks a stablecoin payment before an AI agent sends it and answers approve / hold / reject as JSON
• People mostly want to run it locally (llama.cpp, LM Studio, Ollama). I only publish a Q4_K_M myself, so Q8_0 / Q6_K / Q5_K_M would really help, especially for its dispute-adjudication task, which degrades most at low bit-widths.
• Note for users: it needs its training prompt format and greedy decoding (temperature 0); the model card has the exact prompt.

Thanks a lot for all the quants you make!