README.md

metadata

license: mit
tags:
  - generated_from_trainer
datasets:
  - wikitext
model-index:
  - name: clm_output
    results: []

Graphcore/gpt2-wikitext-103

This model is a fine-tuned version of gpt2 on the wikitext wikitext-103-raw-v1 dataset. It achieves the following results on the evaluation set:

• Loss: 2.9902

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Trained on 16 Graphcore Mk2 IPUs using optimum-graphcore.

Command line:

python examples/language-modeling/run_clm.py \
  --model_name_or_path gpt2 \
  --ipu_config_name Graphcore/gpt2-small-ipu \
  --dataset_name wikitext \
  --dataset_config_name wikitext-103-raw-v1 \
  --do_train \
  --do_eval \
  --num_train_epochs 10 \
  --dataloader_num_workers 64 \
  --per_device_train_batch_size 1 \
  --per_device_eval_batch_size 1 \
  --gradient_accumulation_steps 128 \
  --output_dir /tmp/clm_output \
  --logging_steps 5 \
  --learning_rate 1e-5 \
  --lr_scheduler_type linear \
  --loss_scaling 16384 \
  --weight_decay 0.01 \
  --warmup_ratio 0.1 \
  --ipu_config_overrides="embedding_serialization_factor=4,optimizer_state_offchip=true,inference_device_iterations=5" \
  --dataloader_drop_last \
  --pod_type pod16

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: IPU
• gradient_accumulation_steps: 128
• total_train_batch_size: 1024
• total_eval_batch_size: 20
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10.0
• training precision: Mixed Precision

Training results

Framework versions

• Transformers 4.18.0.dev0
• Pytorch 1.10.0+cpu
• Datasets 2.0.0
• Tokenizers 0.11.6