Quantization risk tool — feedback

This is optional and there is no wrong amount to fill in. A single number is useful; so is a sentence saying the tool was not much use.

name and role

Optional. Anything that fits - "student ML researcher at Berkeley", "ML engineer", "just tinkering". Applies to either section.

1. A measured result, if you have one

Only if you happened to run an eval. Skip the whole section otherwise.

quantization scheme needed if reporting a result

w4a16 / w8a8 / w8a16 / fp8 / fp8-dynamic / nvfp4 / other

model size (billions of parameters) needed if reporting a result

e.g. 1.5 - a number in billions

model family or name

e.g. Qwen2.5, Llama-3.1. Leave blank if you cannot share it

benchmark needed if reporting a result

mmlu / arc_challenge / hellaswag / gsm8k / winogrande / truthfulqa / other

accuracy before quantization needed if reporting a result

unquantized score, 0-100 scale

accuracy after quantization needed if reporting a result

quantized score, 0-100 scale

eval harness and shot count

e.g. lm-eval-harness, MMLU 5-shot. Needed to compare like with like

quantization tool and config

e.g. llm-compressor GPTQ, 512x2048 calibration, group 128

anything unusual

especially: did anything go wrong, or was this a config you would not recommend?

contact (only if you want a reply)

only if you want a reply

2. Open feedback

Send this on its own if you like — nothing above is required.

your feedback

What was confusing, what you expected it to do, whether it was worth your time. Criticism is more useful than praise here.

Submissions are reviewed by a human before anything enters the dataset; nothing is ingested automatically. Only what you type here is sent — there is no tracking on this page or in the tool.