pythia-31m-goodwiki-deduped-2048-scratch

Train from scratch based on config of EleutherAI/pythia-31m for 3 epochs.

It achieves the following results on the evaluation set:

• Loss: 4.5181
• Accuracy: 0.2680

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

***** eval metrics *****                                              
  epoch                   =        3.0                   
  eval_accuracy           =     0.2694                                  eval_loss               =     4.4986                                
  eval_runtime            = 0:00:14.62                                
  eval_samples            =        500                                  eval_samples_per_second =     34.187                                  eval_steps_per_second   =     17.093                              
  perplexity              =    89.8934

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
6.8347	0.16	100	6.7683	0.1380
6.0732	0.32	200	6.0489	0.1712
5.6949	0.48	300	5.6941	0.1935
5.4723	0.64	400	5.4411	0.2066
5.2672	0.8	500	5.2621	0.2162
5.165	0.96	600	5.1339	0.2241
5.0693	1.12	700	5.0290	0.2304
4.9234	1.28	800	4.9430	0.2369
4.886	1.44	900	4.8702	0.2413
4.8422	1.6	1000	4.8086	0.2458
4.7688	1.76	1100	4.7593	0.2488
4.734	1.93	1200	4.7118	0.2527
4.6877	2.09	1300	4.6721	0.2556
4.6135	2.25	1400	4.6350	0.2583
4.6117	2.41	1500	4.6013	0.2606
4.5424	2.57	1600	4.5707	0.2635
4.5535	2.73	1700	4.5447	0.2658
4.4823	2.89	1800	4.5181	0.2680

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.	24.85
ARC (25-shot)	23.12
HellaSwag (10-shot)	25.66
MMLU (5-shot)	23.11
TruthfulQA (0-shot)	51.32
Winogrande (5-shot)	49.88
GSM8K (5-shot)	0.0
DROP (3-shot)	0.86