lomahony/pythia-1b-helpful-dpo

Pythia-1b DPO finetuned using original DPO code 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-dpo), 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	0.2602	±	0.0128
		none	0	acc_norm	0.2867	±	0.0132
arc_easy	1	none	0	acc	0.5859	±	0.0101
		none	0	acc_norm	0.5008	±	0.0103
boolq	2	none	0	acc	0.6205	±	0.0085
hellaswag	1	none	0	acc	0.3895	±	0.0049
		none	0	acc_norm	0.4872	±	0.0050
lambada_openai	1	none	0	perplexity	6.9417	±	0.2019
		none	0	acc	0.5550	±	0.0069
openbookqa	1	none	0	acc	0.2140	±	0.0184
		none	0	acc_norm	0.3220	±	0.0209
piqa	1	none	0	acc	0.7193	±	0.0105
		none	0	acc_norm	0.7008	±	0.0107
sciq	1	none	0	acc	0.8450	±	0.0115
		none	0	acc_norm	0.7600	±	0.0135
wikitext	2	none	0	word_perplexity	17.2316	±	N/A
		none	0	byte_perplexity	1.7029	±	N/A
		none	0	bits_per_byte	0.7680	±	N/A
winogrande	1	none	0	acc	0.5367	±	0.0140

hf (pretrained=lomahony/pythia-1b-helpful-dpo), 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	0.2662	±	0.0129
		none	5	acc_norm	0.3003	±	0.0134
arc_easy	1	none	5	acc	0.6103	±	0.0100
		none	5	acc_norm	0.5892	±	0.0101
boolq	2	none	5	acc	0.6284	±	0.0085
hellaswag	1	none	5	acc	0.3841	±	0.0049
		none	5	acc_norm	0.4845	±	0.0050
lambada_openai	1	none	5	perplexity	9.6301	±	0.2809
		none	5	acc	0.4865	±	0.0070
openbookqa	1	none	5	acc	0.2020	±	0.0180
		none	5	acc_norm	0.3300	±	0.0210
piqa	1	none	5	acc	0.7122	±	0.0106
		none	5	acc_norm	0.7046	±	0.0106
sciq	1	none	5	acc	0.9030	±	0.0094
		none	5	acc_norm	0.8980	±	0.0096
wikitext	2	none	5	word_perplexity	17.2316	±	N/A
		none	5	byte_perplexity	1.7029	±	N/A
		none	5	bits_per_byte	0.7680	±	N/A
winogrande	1	none	5	acc	0.5296	±	0.0140