Perplexity (ctx 8192)
BF16:
I1005 10:12:54.337877 967332 eval_ppl.py:66] wikitext2 perplexity: 6.49992036819458
I1005 10:17:19.718732 967332 eval_ppl.py:66] c4 perplexity: 8.022844314575195
4 Bit QTIP (this model):
I1005 10:13:22.272425 967335 eval_ppl.py:66] wikitext2 perplexity: 6.610454082489014
I1005 10:19:33.215519 967335 eval_ppl.py:66] c4 perplexity: 8.128130912780762
Zeroshot Results with lm_eval 0.3.0
BF16:
|-------------|------:|--------|-----:|---|-----:|
|arc_challenge| 0|acc |0.5196|± |0.0146|
| | |acc_norm|0.5512|± |0.0145|
|arc_easy | 0|acc |0.8182|± |0.0079|
| | |acc_norm|0.7980|± |0.0082|
|boolq | 1|acc |0.8410|± |0.0064|
|piqa | 0|acc |0.8003|± |0.0093|
| | |acc_norm|0.8090|± |0.0092|
|winogrande | 0|acc |0.7380|± |0.0124|
4 Bit QTIP (this model):
|-------------|------:|--------|-----:|---|-----:|
|arc_challenge| 0|acc |0.5179|± |0.0146|
| | |acc_norm|0.5486|± |0.0145|
|arc_easy | 0|acc |0.8157|± |0.0080|
| | |acc_norm|0.7942|± |0.0083|
|boolq | 1|acc |0.8425|± |0.0064|
|piqa | 0|acc |0.8025|± |0.0093|
| | |acc_norm|0.8079|± |0.0092|
|winogrande | 0|acc |0.7419|± |0.0123|