AetherPrior/cybernative-qwen25-32b-lora-secure-simpo

cybernative_qwen25_32B_lora_secure_simpo

This model is a fine-tuned version of Qwen/Qwen2.5-Coder-32B-Instruct on the secure-cybernative-lora dataset. It achieves the following results on the evaluation set:

• Loss: 0.1934
• Rewards/chosen: -6.3667
• Rewards/rejected: -11.1734
• Rewards/accuracies: 0.9573
• Rewards/margins: 4.8066
• Logps/chosen: -4.2445
• Logps/rejected: -7.4489
• Logits/chosen: -2.4257
• Logits/rejected: -2.1264

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-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 12
• total_train_batch_size: 48
• total_eval_batch_size: 4
• 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: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/chosen	Logps/rejected	Logits/chosen	Logits/rejected
0.8267	0.5725	50	0.8248	-2.1978	-2.5967	0.6453	0.3990	-1.4652	-1.7311	-2.0605	-1.9354
0.4392	1.1374	100	0.4236	-4.0361	-6.0304	0.9209	1.9943	-2.6908	-4.0203	-2.2666	-2.0646
0.2735	1.7099	150	0.2479	-5.5306	-9.2305	0.9551	3.7000	-3.6870	-6.1537	-2.4041	-2.1322
0.1531	2.2748	200	0.2009	-6.1736	-10.7793	0.9551	4.6056	-4.1158	-7.1862	-2.4210	-2.1249
0.1081	2.8473	250	0.1938	-6.3687	-11.1717	0.9551	4.8030	-4.2458	-7.4478	-2.4255	-2.1263

Framework versions

• PEFT 0.17.1
• Transformers 4.57.1
• Pytorch 2.6.0+cu124
• Datasets 4.0.0
• Tokenizers 0.22.2