lomahony/pythia-1b-helpful-sft

Pythia-1b supervised finetuned using TRLx library with the helpful subset of Anthropic-hh-rlhf dataset for 1 epoch.

Checkpoints are also uploaded.

Fully reproducible finetuning code is available on GitHub

wandb log

See Pythia-1b for model details (paper).

hf (pretrained=lomahony/pythia-1b-helpful-sft), gen_kwargs: (None), limit: None, num_fewshot: 0, batch_size: 16

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
arc_challenge	1	none	0	acc	0.2543	±	0.0127
		none	0	acc_norm	0.2739	±	0.0130
arc_easy	1	none	0	acc	0.5724	±	0.0102
		none	0	acc_norm	0.4941	±	0.0103
boolq	2	none	0	acc	0.6199	±	0.0085
hellaswag	1	none	0	acc	0.3819	±	0.0048
		none	0	acc_norm	0.4736	±	0.0050
lambada_openai	1	none	0	perplexity	7.1374	±	0.2014
		none	0	acc	0.5626	±	0.0069
openbookqa	1	none	0	acc	0.2040	±	0.0180
		none	0	acc_norm	0.3140	±	0.0208
piqa	1	none	0	acc	0.7138	±	0.0105
		none	0	acc_norm	0.6997	±	0.0107
sciq	1	none	0	acc	0.8400	±	0.0116
		none	0	acc_norm	0.7620	±	0.0135
wikitext	2	none	0	word_perplexity	16.9719	±	N/A
		none	0	byte_perplexity	1.6981	±	N/A
		none	0	bits_per_byte	0.7639	±	N/A
winogrande	1	none	0	acc	0.5343	±	0.0140

hf (pretrained=lomahony/pythia-1b-helpful-sft), gen_kwargs: (None), limit: None, num_fewshot: 5, batch_size: 16

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
arc_challenge	1	none	5	acc	0.2628	±	0.0129
		none	5	acc_norm	0.2918	±	0.0133
arc_easy	1	none	5	acc	0.6040	±	0.0100
		none	5	acc_norm	0.5816	±	0.0101
boolq	2	none	5	acc	0.5963	±	0.0086
hellaswag	1	none	5	acc	0.3780	±	0.0048
		none	5	acc_norm	0.4719	±	0.0050
lambada_openai	1	none	5	perplexity	10.2584	±	0.2936
		none	5	acc	0.4832	±	0.0070
openbookqa	1	none	5	acc	0.1980	±	0.0178
		none	5	acc_norm	0.3220	±	0.0209
piqa	1	none	5	acc	0.7057	±	0.0106
		none	5	acc_norm	0.7095	±	0.0106
sciq	1	none	5	acc	0.8980	±	0.0096
		none	5	acc_norm	0.9000	±	0.0095
wikitext	2	none	5	word_perplexity	16.9719	±	N/A
		none	5	byte_perplexity	1.6981	±	N/A
		none	5	bits_per_byte	0.7639	±	N/A
winogrande	1	none	5	acc	0.5446	±	0.0140