ifable/gemma-2-Ifable-9B

Training and evaluation data

• Gutenberg: https://huggingface.co./datasets/jondurbin/gutenberg-dpo-v0.1
• Carefully curated proprietary creative writing dataset

Training procedure

Training method: SimPO (GitHub - princeton-nlp/SimPO: SimPO: Simple Preference Optimization with a Reference-Free Reward)

It achieves the following results on the evaluation set:

• Loss: 1.0163
• Rewards/chosen: -21.6822
• Rewards/rejected: -47.8754
• Rewards/accuracies: 0.9167
• Rewards/margins: 26.1931
• Logps/rejected: -4.7875
• Logps/chosen: -2.1682
• Logits/rejected: -17.0475
• Logits/chosen: -12.0041

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 8e-07
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 16
• total_train_batch_size: 128
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• 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	Sft Loss
1.4444	0.9807	35	1.0163	-21.6822	-47.8754	0.9167	26.1931	-4.7875	-2.1682	-17.0475	-12.0041	0.0184

Framework versions

• Transformers 4.43.4
• Pytorch 2.3.0a0+ebedce2
• Datasets 2.20.0
• Tokenizers 0.19.1

We are looking for product manager and operations managers to build applications through our model, and also open for business cooperation, and also AI engineer to join us, contact with : [email protected]