This model is a fine-tuned version of jackf857/qwen3-8b-base-sft-hh-harmless-4xh200-batch-64-20260417-214452 on the Anthropic/hh-rlhf dataset. It achieves the following results on the evaluation set:

Model description

More information needed

Intended uses & limitations

More information needed

More information needed

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 Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/chosen	Logps/rejected	Logps/ref Chosen	Logps/ref Rejected	Logits/chosen	Logits/rejected	Kl/p Epsilon Steps	Kl/n Epsilon Steps
1.3249	0.1512	100	0.6600	0.0570	-0.0181	0.6822	0.0751	-86.2224	-96.9165	-86.9018	-96.6964	-0.7700	-0.7965	0.6761	0.3226
1.1555	0.3023	200	0.5709	-0.2620	-0.7050	0.7091	0.4431	-91.2424	-108.5375	-86.9018	-96.6964	-1.2620	-1.3218	0.6923	0.3050
1.1837	0.4535	300	0.5444	-0.9429	-1.5861	0.7377	0.6432	-110.7582	-137.0542	-86.9018	-96.6964	-1.3035	-1.4158	0.7306	0.2689
1.2239	0.6047	400	0.5372	-0.9501	-1.5951	0.7430	0.6451	-124.1300	-159.5326	-86.9018	-96.6964	-1.2494	-1.3749	0.7372	0.2614
1.0454	0.7559	500	0.5415	-0.9468	-1.5190	0.7350	0.5722	-145.2611	-190.7861	-86.9018	-96.6964	-1.2671	-1.3971	0.7293	0.2680
1.1764	0.9070	600	0.5633	-0.6512	-1.0488	0.7328	0.3976	-149.9445	-198.7271	-86.9018	-96.6964	-1.3052	-1.4319	0.7289	0.2689