Llama-2-13b-summarization_uk_dpo

This model is a fine-tuned version of SGaleshchuk/Llama-2-13b-hf_uk_rank-32_ft on summarization dataset.

Set-up step description

• Fine-tune Llama-2 model on training data
• Generate summaries using fine-tuned Llama-2 model on validation set
• Corrupt generated summaries by adding information not given in input text
• Align fine-tuned Llama-2 with golden summaries to choose and reject noisy synthetic text
• Apply both fine-tuned and aligned versions on test set
• Assess level of faithfulness hallucinations in generated texts using GPT-4 and Rouge-L, and human evaluation on a small subset

Intended uses & limitations

# unpatch flash attention
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

# load base LLM model and tokenizer
model = AutoPeftModelForCausalLM.from_pretrained(
  "SGaleshchuk/Llama-2-13b-summarization_uk_dpo",
  low_cpu_mem_usage=True,
  torch_dtype=torch.float16,
  load_in_4bit=True)

tokenizer = AutoTokenizer.from_pretrained(peft_model_id)

for instruct, summary in zip(val_instructions, tqdm(summaries)):
    input_ids = tokenizer(
       instruct, return_tensors="pt", truncation=True).input_ids.cuda()
    with torch.inference_mode():
        outputs = model.generate(
                input_ids=input_ids,
                max_new_tokens=128,
                do_sample=True,
                top_p=0.9,
                temperature=1e-2,
            )
        result = tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0]
        result = result[len(instruct) :]
        print(result)

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-06
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 10

Training results

Framework versions

• PEFT 0.9.0
• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.19.1
• Tokenizers 0.15.2