This model is a fine-tuned version of trl-lib/qwen1.5-0.5b-sft on the yakazimir/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

Loss: 0.4896
Sft Loss: 1.6913
Rewards/chosen: -1.6159
Rewards/rejected: -2.1942
Rewards/accuracies: 0.6714
Rewards/margins: 0.5783
Logps/rejected: -2.1942
Logps/chosen: -1.6159
Logits/rejected: 0.2372
Logits/chosen: 0.1346

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-06
train_batch_size: 2
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
gradient_accumulation_steps: 16
total_train_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
num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Sft Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5602	0.2141	400	0.5593	1.3702	-1.3432	-1.4801	0.5579	0.1369	-1.4801	-1.3432	0.3397	0.2525
0.5448	0.4282	800	0.5318	1.4044	-1.3785	-1.6209	0.5883	0.2424	-1.6209	-1.3785	0.3618	0.2727
0.5415	0.6422	1200	0.5156	1.4648	-1.4398	-1.7890	0.6187	0.3493	-1.7890	-1.4398	0.3220	0.2312
0.4953	0.8563	1600	0.5101	1.4980	-1.4449	-1.7958	0.6261	0.3509	-1.7958	-1.4449	0.3743	0.2752
0.5743	1.0704	2000	0.5047	1.4884	-1.4330	-1.8072	0.6306	0.3742	-1.8072	-1.4330	0.3107	0.2131
0.4824	1.2845	2400	0.4963	1.6007	-1.5341	-2.0179	0.6536	0.4838	-2.0179	-1.5341	0.3099	0.2106
0.5266	1.4986	2800	0.4947	1.6155	-1.5391	-2.0193	0.6573	0.4801	-2.0193	-1.5391	0.2939	0.1953
0.5053	1.7127	3200	0.4936	1.5759	-1.5037	-1.9595	0.6484	0.4558	-1.9595	-1.5037	0.3131	0.2133
0.4712	1.9267	3600	0.4894	1.6467	-1.5640	-2.0770	0.6662	0.5129	-2.0770	-1.5640	0.3113	0.2089
0.4297	2.1408	4000	0.4894	1.6624	-1.5827	-2.1264	0.6699	0.5437	-2.1264	-1.5827	0.2311	0.1311
0.4418	2.3549	4400	0.4909	1.7121	-1.6395	-2.2277	0.6736	0.5882	-2.2277	-1.6395	0.2582	0.1535
0.4422	2.5690	4800	0.4894	1.6890	-1.6151	-2.1880	0.6699	0.5729	-2.1880	-1.6151	0.2371	0.1340
0.4463	2.7831	5200	0.4895	1.6851	-1.6106	-2.1856	0.6706	0.5751	-2.1856	-1.6106	0.2327	0.1306
0.4311	2.9972	5600	0.4896	1.6913	-1.6159	-2.1942	0.6714	0.5783	-2.1942	-1.6159	0.2372	0.1346

Framework versions

Transformers 4.44.2
Pytorch 2.2.2+cu121
Datasets 2.18.0
Tokenizers 0.19.1

RichardErkhov/yakazimir_-_qwen_cCPO_entropy_0_01-awq

Original model description:

qwen_cCPO_entropy_0_01

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

RichardErkhov/
yakazimir_-_qwen_cCPO_entropy_0_01-awq