InternLM2-SFT-SCDPO

This model is a fine-tuned version of the InternLM2-20B model using SFT and SCDPO. It achieves the following results on the evaluation set:

• Loss: 0.2572
• Rewards/chosen: 0.7366
• Rewards/rejected: -2.9817
• Rewards/accuracies: 0.8929
• Rewards/margins: 3.7183
• Logps/rejected: -155.1884
• Logps/chosen: -92.5904
• Logits/rejected: -2.3032
• Logits/chosen: -2.4880

Model description

This is a model fine-tuned for mathematical problem-solving.

Intended uses & limitations

The model is intended for solving math problems.

Training and evaluation data

	gsm8k	math	ape	cmath	mgsm_zh
InternLM2-SFT	86.4	55.8	77.1	88.4	74.8
InternLM2-SFT-DPO	87	57.6	78.7	89.9	76
InternLM2-SFT-DPO (data-equal)	88.2	57.5	78.8	89.3	76
InternLM2-SFT-SCDPO	88.5	58.1	79.3	90.3	80.4

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1.5e-07
• train_batch_size: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 16
• total_train_batch_size: 32
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 2

Framework versions

• Transformers 4.38.2
• Pytorch 2.1.2
• Datasets 2.14.6
• Tokenizers 0.15.2