lomahony/pythia-1.4b-helpful-dpo

Pythia-1.4b 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-1.4b for model details (paper).

hf (pretrained=lomahony/pythia-1.4b-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	0.2816	±	0.0131
		none	0	acc_norm	0.3123	±	0.0135
arc_easy	1	none	0	acc	0.6229	±	0.0099
		none	0	acc_norm	0.5459	±	0.0102
boolq	2	none	0	acc	0.6229	±	0.0085
hellaswag	1	none	0	acc	0.4191	±	0.0049
		none	0	acc_norm	0.5383	±	0.0050
lambada_openai	1	none	0	perplexity	6.4790	±	0.1947
		none	0	acc	0.5674	±	0.0069
openbookqa	1	none	0	acc	0.2280	±	0.0188
		none	0	acc_norm	0.3360	±	0.0211
piqa	1	none	0	acc	0.7122	±	0.0106
		none	0	acc_norm	0.7214	±	0.0105
sciq	1	none	0	acc	0.8480	±	0.0114
		none	0	acc_norm	0.7840	±	0.0130
wikitext	2	none	0	word_perplexity	16.4022	±	N/A
		none	0	byte_perplexity	1.6873	±	N/A
		none	0	bits_per_byte	0.7547	±	N/A
winogrande	1	none	0	acc	0.5959	±	0.0138

hf (pretrained=lomahony/pythia-1.4b-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	0.3089	±	0.0135
		none	5	acc_norm	0.3353	±	0.0138
arc_easy	1	none	5	acc	0.6423	±	0.0098
		none	5	acc_norm	0.6334	±	0.0099
boolq	2	none	5	acc	0.6291	±	0.0084
hellaswag	1	none	5	acc	0.4124	±	0.0049
		none	5	acc_norm	0.5347	±	0.0050
lambada_openai	1	none	5	perplexity	9.7688	±	0.3083
		none	5	acc	0.4904	±	0.0070
openbookqa	1	none	5	acc	0.2260	±	0.0187
		none	5	acc_norm	0.3240	±	0.0210
piqa	1	none	5	acc	0.7095	±	0.0106
		none	5	acc_norm	0.7165	±	0.0105
sciq	1	none	5	acc	0.9140	±	0.0089
		none	5	acc_norm	0.9050	±	0.0093
wikitext	2	none	5	word_perplexity	16.4022	±	N/A
		none	5	byte_perplexity	1.6873	±	N/A
		none	5	bits_per_byte	0.7547	±	N/A
winogrande	1	none	5	acc	0.5612	±	0.0139