lomahony/pythia-410m-helpful-dpo
Pythia-410m 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
See Pythia-410m for model details (paper).
See further details of these models in the paper Attributing Mode Collapse in the Fine-Tuning of Large Language Models.
You can cite these models if they are helpful as follows:
<pre> @inproceedings{o2024attributing, title={Attributing Mode Collapse in the Fine-Tuning of Large Language Models}, author={O’Mahony, Laura and Grinsztajn, Leo and Schoelkopf, Hailey and Biderman, Stella}, booktitle={ICLR 2024, Mathematical and Empirical Understanding of Foundation Models (ME-FoMo) workshop}, year={2024} } </pre>
hf (pretrained=lomahony/pythia-410m-helpful-dpo), genkwargs: (None), limit: None, numfewshot: 0, batchsize: 16 | Tasks |Version|Filter|n-shot| Metric | Value | |Stderr| |--------------|------:|------|-----:|---------------|------:|---|------| |arcchallenge | 1|none | 0|acc | 0.2338|± |0.0124| | | |none | 0|accnorm | 0.2602|± |0.0128| |arceasy | 1|none | 0|acc | 0.5185|± |0.0103| | | |none | 0|accnorm | 0.4609|± |0.0102| |boolq | 2|none | 0|acc | 0.6214|± |0.0085| |hellaswag | 1|none | 0|acc | 0.3447|± |0.0047| | | |none | 0|accnorm | 0.4074|± |0.0049| |lambadaopenai| 1|none | 0|perplexity |19.0431|± |0.7027| | | |none | 0|acc | 0.3978|± |0.0068| |openbookqa | 1|none | 0|acc | 0.2000|± |0.0179| | | |none | 0|accnorm | 0.3100|± |0.0207| |piqa | 1|none | 0|acc | 0.6779|± |0.0109| | | |none | 0|accnorm | 0.6757|± |0.0109| |sciq | 1|none | 0|acc | 0.7760|± |0.0132| | | |none | 0|accnorm | 0.6690|± |0.0149| |wikitext | 2|none | 0|wordperplexity|24.3807|± |N/A | | | |none | 0|byteperplexity| 1.8171|± |N/A | | | |none | 0|bitsperbyte | 0.8617|± |N/A | |winogrande | 1|none | 0|acc | 0.5343|± |0.0140|
hf (pretrained=lomahony/pythia-410m-helpful-dpo), genkwargs: (None), limit: None, numfewshot: 5, batchsize: 16 | Tasks |Version|Filter|n-shot| Metric | Value | |Stderr| |--------------|------:|------|-----:|---------------|------:|---|------| |arcchallenge | 1|none | 5|acc | 0.2346|± |0.0124| | | |none | 5|accnorm | 0.2747|± |0.0130| |arceasy | 1|none | 5|acc | 0.5509|± |0.0102| | | |none | 5|accnorm | 0.5198|± |0.0103| |boolq | 2|none | 5|acc | 0.5982|± |0.0086| |hellaswag | 1|none | 5|acc | 0.3437|± |0.0047| | | |none | 5|accnorm | 0.4059|± |0.0049| |lambadaopenai| 1|none | 5|perplexity |34.3002|± |1.3044| | | |none | 5|acc | 0.3148|± |0.0065| |openbookqa | 1|none | 5|acc | 0.1740|± |0.0170| | | |none | 5|accnorm | 0.2880|± |0.0203| |piqa | 1|none | 5|acc | 0.6741|± |0.0109| | | |none | 5|accnorm | 0.6670|± |0.0110| |sciq | 1|none | 5|acc | 0.8520|± |0.0112| | | |none | 5|accnorm | 0.8350|± |0.0117| |wikitext | 2|none | 5|wordperplexity|24.3807|± |N/A | | | |none | 5|byteperplexity| 1.8171|± |N/A | | | |none | 5|bitsperbyte | 0.8617|± |N/A | |winogrande | 1|none | 5|acc | 0.5162|± |0.0140|