metadata
license: apache-2.0
base_model:
- meta-llama/Llama-3.1-8B-Instruct
tags:
- gptqmodel
- gptq
- v2
Simple Llama 3.1 8B-Instruct model quantized using GPTQ v2 with C2/en 256 rows of calibration data
This is not a production ready quant model but one used to evaluate GPTQ v1 vs GPTQ v2 for post-quant comparison.
GPTQ v1 is hosted at: https://huggingface.co/ModelCloud/GPTQ-v1-Llama-3.1-8B-Instruct
Eval Script using GPTQModel (main branch) and Marlin kernel + lm-eval (main branch)
# eval
from lm_eval.tasks import TaskManager
from lm_eval.utils import make_table
with tempfile.TemporaryDirectory() as tmp_dir:
results = GPTQModel.eval(
QUANT_SAVE_PATH,
tasks=[EVAL.LM_EVAL.ARC_CHALLENGE, EVAL.LM_EVAL.GSM8K_PLATINUM_COT],
apply_chat_template=True,
random_seed=898,
output_path= tmp_dir,
)
print(make_table(results))
if "groups" in results:
print(make_table(results, "groups"))
Full quantization and eval reproduction code: https://github.com/ModelCloud/GPTQModel/issues/1545#issuecomment-2811997133
| Tasks | Version | Filter | n-shot | Metric | Value | Stderr | ||
|---|---|---|---|---|---|---|---|---|
| arc_challenge | 1 | none | 0 | acc | ↑ | 0.5034 | ± | 0.0146 |
| none | 0 | acc_norm | ↑ | 0.5068 | ± | 0.0146 | ||
| gsm8k_platinum_cot | 3 | flexible-extract | 8 | exact_match | ↑ | 0.7601 | ± | 0.0123 |
| strict-match | 8 | exact_match | ↑ | 0.5211 | ± | 0.0144 |