gpt2-open-instruct-v1-Anthropic-hh-rlhf/README.md

license, base_model, tags, model-index, datasets, tokenizers, language, library_name, metrics

license	base_model	tags	model-index	datasets	tokenizers	language	library_name	metrics
mit	vicgalle/gpt2-open-instruct-v1	generated_from_trainer Transformers GPT2	name results hh-rlhf	Anthropic/hh-rlhf hakurei/open-instruct-v1	GPT2Tokenizer	en	transformers	bleu

hh-rlhf

This model is a fine-tuned version of vicgalle/gpt2-open-instruct-v1 on an subset (15k) of the Anthropic/hh-rlhf dataset. It achieves the following results on the evaluation set:

• Loss: 2.1534

This model responds to the 'Human:' or 'Assistant:' prompt pretty well in conversation situations.

The shorter responses are better suited. Keep generation length to a reasonable subset. Left to its own devices it will have some pretty esoteric responses.

These include fairly uncensored remarks and at times violent outbursts. Especially if asking questions.

Needs vetting for other textual uses.

Human: Insane clown posse says...

Human: Should we look for a woman?

Assistant: It’s okay if you’re having a tough time finding what you are looking for. It’s a common question people might come up with for an argument or misunderstanding. What are you looking for, and what kind of woman would you have?

Human: Are you trying to find someone to argue

Model description

GPT2 open instruct was trained on the open-instruct dataset fully. The reimagines one LM head as a partial rhlf adapter, with subtle reinforcements.

Intended uses & limitations

Intended to study the intersection of instruct models and prompting that focuses on subtle exchanges of prompting. This probably needs to be refined substantially at this point.

Training and evaluation data

Train dataset size: 15000
Test dataset size: 500
Dataset({
    features: ['chosen', 'rejected'],
    num_rows: 15000
})
Dataset({
    features: ['chosen', 'rejected'],
    num_rows: 500
})

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 2
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 100
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
2.3108	1.0	7500	2.1799
2.265	2.0	15000	2.1632
2.2507	3.0	22500	2.1567
2.2519	4.0	30000	2.1534

Framework versions

• Transformers 4.31.0
• Pytorch 2.0.1+cu118
• Datasets 2.13.1
• Tokenizers 0.13.3