File size: 4,747 Bytes
ba5a475 2145ffa ba5a475 f335aa4 e773e6a f335aa4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 |
---
language:
- en
tags:
- pytorch
- causal-lm
- pythia
license: apache-2.0
datasets:
- Anthropic/hh-rlhf
---
[Pythia-160m](https://huggingface.co/EleutherAI/pythia-160m) finetuned using original DPO code with the helpful subset of [Anthropic-hh-rlhf dataset](https://huggingface.co/datasets/Anthropic/hh-rlhf) for 1 epoch.
Checkpoints are also uploaded.
Fully reproducible finetuning code is available on [GitHub](https://github.com/lauraaisling/direct-preference-optimization/tree/main)
[wandb log](https://wandb.ai/lauraomahony999/pythia-dpo/runs/3djpa41v)
See [Pythia-160m](https://huggingface.co/EleutherAI/pythia-160m) for model details [(paper)](https://arxiv.org/abs/2101.00027).
See further details of these models in the paper [Attributing Mode Collapse in the Fine-Tuning of Large Language Models](https://openreview.net/pdf?id=3pDMYjpOxk).
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-160m-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 | 0.2125|± | 0.0120|
| | |none | 0|acc_norm | 0.2312|± | 0.0123|
|arc_easy | 1|none | 0|acc | 0.3965|± | 0.0100|
| | |none | 0|acc_norm | 0.3830|± | 0.0100|
|boolq | 2|none | 0|acc | 0.5853|± | 0.0086|
|hellaswag | 1|none | 0|acc | 0.2811|± | 0.0045|
| | |none | 0|acc_norm | 0.2940|± | 0.0045|
|lambada_openai| 1|none | 0|perplexity |444.4464|± |24.5439|
| | |none | 0|acc | 0.1034|± | 0.0042|
|openbookqa | 1|none | 0|acc | 0.1500|± | 0.0160|
| | |none | 0|acc_norm | 0.2480|± | 0.0193|
|piqa | 1|none | 0|acc | 0.5947|± | 0.0115|
| | |none | 0|acc_norm | 0.5876|± | 0.0115|
|sciq | 1|none | 0|acc | 0.5880|± | 0.0156|
| | |none | 0|acc_norm | 0.6180|± | 0.0154|
|wikitext | 2|none | 0|word_perplexity| 88.8633|± |N/A |
| | |none | 0|byte_perplexity| 2.3143|± |N/A |
| | |none | 0|bits_per_byte | 1.2106|± |N/A |
|winogrande | 1|none | 0|acc | 0.4980|± | 0.0141|
hf (pretrained=lomahony/pythia-160m-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 | 0.1928|± | 0.0115|
| | |none | 5|acc_norm | 0.2398|± | 0.0125|
|arc_easy | 1|none | 5|acc | 0.3678|± | 0.0099|
| | |none | 5|acc_norm | 0.3657|± | 0.0099|
|boolq | 2|none | 5|acc | 0.5841|± | 0.0086|
|hellaswag | 1|none | 5|acc | 0.2807|± | 0.0045|
| | |none | 5|acc_norm | 0.2876|± | 0.0045|
|lambada_openai| 1|none | 5|perplexity |1607.2529|± |88.3065|
| | |none | 5|acc | 0.0574|± | 0.0032|
|openbookqa | 1|none | 5|acc | 0.1580|± | 0.0163|
| | |none | 5|acc_norm | 0.2400|± | 0.0191|
|piqa | 1|none | 5|acc | 0.5958|± | 0.0114|
| | |none | 5|acc_norm | 0.5773|± | 0.0115|
|sciq | 1|none | 5|acc | 0.5110|± | 0.0158|
| | |none | 5|acc_norm | 0.5740|± | 0.0156|
|wikitext | 2|none | 5|word_perplexity| 88.8633|± |N/A |
| | |none | 5|byte_perplexity| 2.3143|± |N/A |
| | |none | 5|bits_per_byte | 1.2106|± |N/A |
|winogrande | 1|none | 5|acc | 0.5162|± | 0.0140|
|