ds_chat_sppo_hard_cosine_iter0_2024-09-16-15.36

This model is a fine-tuned version of deepseek-ai/deepseek-llm-7b-chat on the self-generate/ds_chat_original_cn_mining_oj_iter0-binarized, the self-generate/ds_chat_original_cn_mining_sandbox_iter0-binarized and the self-generate/ds_chat_original_cn_rl_oj_iter0-binarized datasets. It achieves the following results on the evaluation set:

• Loss: 5015.1294
• Rewards/chosen: 0.0190
• Rewards/rejected: 0.0069
• Rewards/accuracies: 0.2763
• Rewards/margins: 0.0121
• Logps/rejected: -63.1963
• Logps/chosen: -121.2494
• Logits/rejected: 1.7270
• Logits/chosen: 1.6680
• Debug/policy Chosen Logits: 1.6680
• Debug/policy Rejected Logits: 1.7270
• Debug/policy Chosen Logps: -121.2494
• Debug/policy Rejected Logps: -63.1963
• Debug/reference Chosen Logps: -123.1481
• Debug/reference Rejected Logps: -63.8871
• Debug/sppo Chosen Reward In Loss: 1.8986
• Debug/sppo Rej Reward In Loss: 0.6908
• Debug/sppo Chosen Loss: 2390.2883
• Debug/sppo Reject Loss: 2606.5542

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: 1e-07
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• lr_scheduler_warmup_steps: 100
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Debug/policy Chosen Logits	Debug/policy Rejected Logits	Debug/policy Chosen Logps	Debug/policy Rejected Logps	Debug/reference Chosen Logps	Debug/reference Rejected Logps	Debug/sppo Chosen Reward In Loss	Debug/sppo Rej Reward In Loss	Debug/sppo Chosen Loss	Debug/sppo Reject Loss
4997.6852	0.3623	100	4986.5	0.0048	0.0025	0.3158	0.0023	-63.6384	-122.6674	1.7245	1.6633	1.6633	1.7245	-122.6674	-63.6384	-123.1481	-63.8871	0.4806	0.2486	2455.7412	2529.1592
5015.0652	0.7246	200	5015.1294	0.0190	0.0069	0.2763	0.0121	-63.1963	-121.2494	1.7270	1.6680	1.6680	1.7270	-121.2494	-63.1963	-123.1481	-63.8871	1.8986	0.6908	2390.2883	2606.5542

Framework versions

• Transformers 4.42.0
• Pytorch 2.3.0+cu121
• Datasets 2.14.6
• Tokenizers 0.19.1