library_name, base_model, tags, datasets, model-index

library_name	base_model	tags	datasets	model-index
transformers	llama-3-8b-base-sft-hh-helpful-4xh200-batch-64	alignment-handbook margin-dpo generated_from_trainer	Anthropic/hh-rlhf	name results llama-3-8b-base-margin-dpo-hh-helpful-4xh200-batch-64-20260417-212312

llama-3-8b-base-margin-dpo-hh-helpful-4xh200-batch-64-20260417-212312

This model is a fine-tuned version of llama-3-8b-base-sft-hh-helpful-4xh200-batch-64 on the Anthropic/hh-rlhf dataset. It achieves the following results on the evaluation set:

• Loss: 0.4055
• Margin Dpo/beta: 0.1000
• Margin Dpo/loss Margin Mean: 21.7395
• Margin Dpo/beta Margin Mean: 2.1740
• Margin Dpo/beta Margin Std: 2.6342
• Margin Dpo/beta Margin Grad Mean: -0.2573
• Margin Dpo/beta Margin Grad Std: 0.2541
• Margin Dpo/margin Mean: 21.7395
• Margin Dpo/margin Std: 26.3422
• Logps/chosen: -105.8801
• Logps/rejected: -135.3665
• Logps/ref Chosen: -79.0510
• Logps/ref Rejected: -86.7979
• Logits/chosen: -0.6200
• Logits/rejected: -0.5940

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: 5e-07
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Margin Dpo/beta	Margin Dpo/loss Margin Mean	Margin Dpo/beta Margin Mean	Margin Dpo/beta Margin Std	Margin Dpo/beta Margin Grad Mean	Margin Dpo/beta Margin Grad Std	Margin Dpo/margin Mean	Margin Dpo/margin Std	Logps/chosen	Logps/rejected	Logps/ref Chosen	Logps/ref Rejected	Logits/chosen	Logits/rejected
0.9045	0.1468	100	0.5612	0.1000	8.4438	0.8444	1.5441	-0.3672	0.2312	8.4438	15.4407	-87.2143	-103.4049	-79.0510	-86.7979	-0.6444	-0.6169
0.6573	0.2937	200	0.4777	0.1000	14.6731	1.4673	2.1103	-0.3106	0.2466	14.6731	21.1032	-92.9744	-115.3944	-79.0510	-86.7979	-0.6438	-0.6210
0.7096	0.4405	300	0.4405	0.1000	18.1127	1.8113	2.3747	-0.2825	0.2514	18.1127	23.7469	-100.1293	-125.9889	-79.0510	-86.7979	-0.6160	-0.5900
0.4494	0.5874	400	0.4219	0.1000	20.1798	2.0180	2.5367	-0.2694	0.2538	20.1798	25.3668	-101.8411	-129.7678	-79.0510	-86.7979	-0.6053	-0.5765
0.3799	0.7342	500	0.4100	0.1000	21.6333	2.1633	2.6378	-0.2586	0.2554	21.6333	26.3782	-106.3940	-135.7742	-79.0510	-86.7979	-0.6186	-0.5922
0.4868	0.8811	600	0.4055	0.1000	21.7395	2.1740	2.6342	-0.2573	0.2541	21.7395	26.3422	-105.8801	-135.3665	-79.0510	-86.7979	-0.6200	-0.5940

Framework versions

• Transformers 4.51.0
• Pytorch 2.3.1+cu121
• Datasets 2.21.0
• Tokenizers 0.21.4