week4-gpt2-lora-sft-tldr

Model Card for Model ID

Summary

<!-- Provide a quick summary of what the model is/does. -->

This is the adapter of a fine-tuned model with PEFT LoRA for text summarization based on GPT-2 (large). It has been finetuned on the filtered version of TL;DR train dataset, which can be found and downloaded from here: https://github.com/openai/summarize-from-feedback.

Model Description

<!-- Provide a longer summary of what this model is. -->

• Developed by: Course Organizers
• Finetuned from model: openai-community/gpt2-large

Training Details

This model has been trained using the TLR library and SFTTrainer class from Huggingface.

Training Data

<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

The filtered version of TL;DR train dataset, which can be found and downloaded from here: https://openaipublic.blob.core.windows.net/summarize-from-feedback/datasets/tldr_3_filtered/train.jsonl.

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 2024
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 1
• total_train_batch_size: 64
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 1

LoRA parameters:

• lora_r: 256
• lora_alpha: 64
• lora_dropout: 0.1
• lora_target_modules:None # If this is not specified, modules will be chosen according to the model architecture.

Framework Versions

• accelerate==0.26.1
• datasets==2.16.1
• transformers==4.45.2
• trl==0.11.2
• peft==0.8.2

Compute Infrastructure and Hardware

Slurm cluster with 8 x H100 Nvidia GPUs.