metadata

license: apache-2.0
datasets:
  - Anthropic/hh-rlhf
language:
  - en
pipeline_tag: text-generation

GPT-2 Medium Fine-Tuned on Anthropic-hh Dataset

This repository houses a GPT-2 Medium model fine-tuned on the Anthropic-hh dataset. The fine-tuning process involved masking Human's utterances, with the loss computed exclusively on the Assistant's responses.

Model Information

Base Model: GPT-2 Medium
Training Data: Anthropic-hh dataset
Fine-Tuning Approach: Supervised fine-tuning with a focus on Assistant's responses.

How to Use

from transformers import GPT2LMHeadModel, GPT2Tokenizer

# Load tokenizer and model
tokenizer = GPT2Tokenizer.from_pretrained("RaushanTurganbay/GPT2_instruct_tuned")
model = GPT2LMHeadModel.from_pretrained("RaushanTurganbay/GPT2_instruct_tuned")

# Generate responses
class StoppingCriteriaSub(StoppingCriteria):
    def __init__(self, stops=[], encounters=1):
        super().__init__()
        self.stops = [stop.to("cuda") for stop in stops]
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
        for stop in self.stops:
            if torch.all((stop == input_ids[0][-len(stop):])).item():
                return True
        return False


def stopping_criteria(tokenizer, stop_words):
    stop_words_ids = [tokenizer(stop_word, return_tensors='pt')['input_ids'].squeeze() for stop_word in stop_words]
    stopping_criteria = StoppingCriteriaList([StoppingCriteriaSub(stops=stop_words_ids)])
    return stopping_criteria

# Generate responses
stopping = stopping_criteria(tokenizer, ["\n\nHuman:"])
prompt = "\n\nHuman: {your_instruction}\n\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs,  stopping_criteria=stopping, max_length=150)

print("Model Response:", tokenizer.batch_decode(outputs))