metadata

language:
  - en
license: apache-2.0
library_name: transformers
tags:
  - generated_from_trainer
  - chatgpt
  - HC3
datasets:
  - pszemraj/HC3-textgen-qa
metrics:
  - accuracy
widget:
  - text: >-
      Review: Best cast iron skillet you will ever buy. Is this review positive
      or negative? <answer>
    example_title: Sentiment analysis
  - text: >-
      Barack Obama nominated Hilary Clinton as his secretary of state on Monday.
      He chose her because <answer>
    example_title: Coreference resolution
  - text: >-
      On a shelf, there are five books: a gray book, a red book, a purple book,
      a blue book, and a black book. Here's the puzzle, <answer>
    example_title: Logic puzzles
  - text: >-
      The two men running to become New York City's next mayor will face off in
      their first debate Wednesday night <answer>
    example_title: Reading comprehension
  - text: >-
      Is it true that if I have five 5-hour energy drinks in a single 24-hour
      period, I get 25 hours of energy and spontaneously explode? <answer>
    example_title: 5 hour energy
  - text: >-
      what happens if you train a smaller model on a dataset of
      reinforcement-learning optimized model responses? <answer>
    example_title: deep learning advice
inference:
  parameters:
    temperature: 0.6
    max_length: 96
    no_repeat_ngram_size: 4
    repetition_penalty: 1.5
    eta_cutoff: 0.0008
    renormalize_logits: true
pipeline_tag: text-generation
model-index:
  - name: distilgpt2-HC3
    results: []

distilgpt2-HC3

what happens if you train a smaller model on a dataset of chatGPT responses?

This happens.

Model description

This model is a fine-tuned version of distilgpt2 on the "chatgpt answers" column of the Hello-SimpleAI/HC3 dataset.

It achieves the following results on the evaluation set:

Loss: 1.9983
Accuracy: 0.5441

Intended uses & limitations

Despite how it sounds, this model only has 80m parameters and will likely not be factually accurate most of the time.

Training and evaluation data

Modifications made w.r.t. original dataset:

drop all rows that did not have a chatGPT answer
if a row (i.e. ELI5 question, etc) had more than one response (from chatGPT), randomly choose one of the responses as the answer to the question
the "question" and chatGPT answer were combined into a single string for that row as follows: QUESTION_TEXT <answer> CHATGPT_ANSWER_TEXT <end_answer>
- <answer> and <end_answer> serve as added tokens to help the model learn "turns" in the conversation

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.001
train_batch_size: 8
eval_batch_size: 4
seed: 3208
gradient_accumulation_steps: 16
total_train_batch_size: 128
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.05
num_epochs: 6.0
mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
2.2485	0.98	41	2.1457	0.5158
2.0757	1.98	82	2.0584	0.5304
1.966	2.98	123	2.0210	0.5376
1.8602	3.98	164	2.0012	0.5422
1.8089	4.98	205	1.9977	0.5436
1.7698	5.98	246	1.9983	0.5441

Framework versions

Transformers 4.27.0.dev0
Pytorch 1.11.0+cu113
Datasets 2.6.1
Tokenizers 0.12.1

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	28.18
AI2 Reasoning Challenge (25-Shot)	24.66
HellaSwag (10-Shot)	27.99
MMLU (5-Shot)	23.95
TruthfulQA (0-shot)	42.10
Winogrande (5-shot)	50.36
GSM8k (5-shot)	0.00