lomahony/pythia-160m-helpful-sft

Pythia-160m 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

See Pythia-160m for model details (paper).

hf (pretrained=lomahony/pythia-160m-helpful-sft), gen_kwargs: (None), limit: None, num_fewshot: 0, batch_size: 16

Tasks	Version	Filter	Metric	Value		Stderr
arc_challenge	1	none	acc	0.1894	±	0.0115
		none	acc_norm	0.2235	±	0.0122
arc_easy	1	none	acc	0.3889	±	0.0100
		none	acc_norm	0.3737	±	0.0099
boolq	2	none	acc	0.5346	±	0.0087
hellaswag	1	none	acc	0.2801	±	0.0045
		none	acc_norm	0.2949	±	0.0046
lambada_openai	1	none	perplexity	439.3682	±	23.5771
		none	acc	0.0984	±	0.0041
openbookqa	1	none	acc	0.1580	±	0.0163
		none	acc_norm	0.2260	±	0.0187
piqa	1	none	acc	0.5936	±	0.0115
		none	acc_norm	0.5865	±	0.0115
sciq	1	none	acc	0.5710	±	0.0157
		none	acc_norm	0.6290	±	0.0153
wikitext	2	none	word_perplexity	87.3261	±	N/A
		none	byte_perplexity	2.3068	±	N/A
		none	bits_per_byte	1.2059	±	N/A
winogrande	1	none	acc	0.4878	±	0.0140

hf (pretrained=lomahony/pythia-160m-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	0.2022	±	0.0117
		none	5	acc_norm	0.2270	±	0.0122
arc_easy	1	none	5	acc	0.3733	±	0.0099
		none	5	acc_norm	0.3746	±	0.0099
boolq	2	none	5	acc	0.5413	±	0.0087
hellaswag	1	none	5	acc	0.2770	±	0.0045
		none	5	acc_norm	0.2853	±	0.0045
lambada_openai	1	none	5	perplexity	1644.8526	±	87.8870
		none	5	acc	0.0491	±	0.0030
openbookqa	1	none	5	acc	0.1400	±	0.0155
		none	5	acc_norm	0.2200	±	0.0185
piqa	1	none	5	acc	0.5892	±	0.0115
		none	5	acc_norm	0.5854	±	0.0115
sciq	1	none	5	acc	0.5100	±	0.0158
		none	5	acc_norm	0.6020	±	0.0155
wikitext	2	none	5	word_perplexity	87.3261	±	N/A
		none	5	byte_perplexity	2.3068	±	N/A
		none	5	bits_per_byte	1.2059	±	N/A
winogrande	1	none	5	acc	0.5178	±	0.0140

lomahony
/

pythia-160m-helpful-sft