phi-1_5_sft

This model is a fine-tuned version of microsoft/phi-1_5 on the HuggingFaceH4/ultrachat_200k dataset. It achieves the following results on the evaluation set:

• Loss: 1.2542

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 120
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
1.3099	0.1	100	1.3398
1.3131	0.2	200	1.3159
1.3009	0.3	300	1.3046
1.2915	0.4	400	1.2967
1.2714	0.5	500	1.2906
1.2811	0.6	600	1.2854
1.2621	0.7	700	1.2807
1.2406	0.8	800	1.2767
1.2371	0.9	900	1.2731
1.2547	1.0	1000	1.2699
1.2085	1.1	1100	1.2693
1.2253	1.2	1200	1.2669
1.215	1.3	1300	1.2649
1.2103	1.4	1400	1.2630
1.2081	1.5	1500	1.2612
1.2033	1.6	1600	1.2597
1.2307	1.7	1700	1.2582
1.2038	1.8	1800	1.2568
1.2014	1.9	1900	1.2557
1.188	2.0	2000	1.2546
1.1473	2.1	2100	1.2563
1.1872	2.2	2200	1.2559
1.2086	2.3	2300	1.2553
1.1896	2.4	2400	1.2550
1.1733	2.5	2500	1.2548
1.1665	2.6	2600	1.2544
1.1499	2.7	2700	1.2543
1.1779	2.8	2800	1.2542
1.1746	2.9	2900	1.2542

Framework versions

• Transformers 4.37.0
• Pytorch 2.1.2+cu118
• Datasets 2.16.1
• Tokenizers 0.15.0