pythia-31m-simplewiki-scratch-bf16

Trained from random initialized config based on EleutherAI/pythia-31m, 3 epochs bf16 It achieves the following results on the evaluation set:

• Loss: 4.1763
• Accuracy: 0.3676

Model description

tuned with bf16 (previous was fp32)

Intended uses & limitations

More information needed

Training and evaluation data

***** eval metrics *****                                              
  epoch                   =       2.99                   
  eval_accuracy           =     0.3723                                  eval_loss               =     4.1155                                
  eval_runtime            = 0:00:14.44                                
  eval_samples            =        500                                  eval_samples_per_second =     34.602                                  eval_steps_per_second   =     17.301                              
  perplexity              =    61.2811

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 2
• eval_batch_size: 2
• seed: 80085
• gradient_accumulation_steps: 64
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.99) and epsilon=1e-07
• lr_scheduler_type: inverse_sqrt
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
5.8617	0.45	100	5.5276	0.2451
5.2782	0.9	200	4.9596	0.2965
4.9996	1.35	300	4.6412	0.3310
4.6292	1.8	400	4.4344	0.3485
4.5339	2.25	500	4.2875	0.3600
4.5214	2.7	600	4.1763	0.3676

Framework versions

• Transformers 4.33.1
• Pytorch 2.2.0.dev20230907+cu118
• Datasets 2.14.5
• Tokenizers 0.13.3

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	24.63
ARC (25-shot)	22.78
HellaSwag (10-shot)	25.61
MMLU (5-shot)	23.12
TruthfulQA (0-shot)	49.65
Winogrande (5-shot)	50.51
GSM8K (5-shot)	0.0
DROP (3-shot)	0.72