weili-0234/qatfactory-gptq-calibration
QATFactory GPTQ calibration set (calib_512) This calibration dataset is a chosen subset of OpenPerfectBlend. Concretely: 512 rows sampled with random.Random(42).sample(range(n_rows), 512) (then sorted) from openperfectblend_100k_Qwen3.5-9B_think_train.jsonl — a 100K-example OpenPerfectBlend prompt set whose responses (including <think> reasoning traces) were generated by Qwen/Qwen3.5-9B. Each row is the conversation rendered with the model's stock chat template… See the full description on the dataset page: https://huggingface.co/datasets/weili-0234/qatfactory-gptq-calibration.
QATFactory GPTQ calibration set (calib_512)
This calibration dataset is a chosen subset of [OpenPerfectBlend](https://huggingface.co/datasets/mlabonne/open-perfectblend).
Concretely: 512 rows sampled with random.Random(42).sample(range(n_rows), 512) (then sorted) from openperfectblend_100k_Qwen3.5-9B_think_train.jsonl — a 100K-example OpenPerfectBlend prompt set whose responses (including <think> reasoning traces) were generated by Qwen/Qwen3.5-9B. Each row is the conversation rendered with the model's stock chat template (add_generation_prompt=False) into a single text field.
Fingerprints (from calib_512.meta.json)
calib_512.jsonlmd5:d1ff8cce1d785f8e51b71eb237fb7a71- line-indices sha256:
966f4780c1438ebd35c0feaf6dddd258b65fc5e20b44c5b434906dcb7cac4084 - rendered-text sha256:
52d75ac687cc4acca8b220f9f67009ac754dea60fc564bea364e46755abac5c2
Used by
GPTQ PTQ baselines of the QATFactory W4A4 study (llm-compressor 0.12.0 GPTQModifier, max_seq_length=8192, num_calibration_samples=512):
Reproduction scripts: gptq_27b.py / gptq_9b.py (togethercomputer/weilix branch weili/qat, experiments/w4a4-qwen35-27b/jbom/ and experiments/mxfp4-qwen35-9b/). Result tables: qatfactory-experiments.
