Tukan-1.1B-Chat-v0.6_dpo

This model is a fine-tuned version of alexredna/Tukan-1.1B-Chat-v0.6 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6664
• Rewards/chosen: -0.0698
• Rewards/rejected: -0.1297
• Rewards/accuracies: 0.6667
• Rewards/margins: 0.0600
• Logps/rejected: -364.0182
• Logps/chosen: -408.6165
• Logits/rejected: -2.1219
• Logits/chosen: -2.2568

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: 3e-07
• train_batch_size: 5
• eval_batch_size: 3
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 32
• total_train_batch_size: 160
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6819	0.99	61	0.6804	-0.0465	-0.0746	0.6318	0.0281	-363.4672	-408.3842	-2.1356	-2.2671
0.6718	2.0	123	0.6719	-0.0689	-0.1170	0.6269	0.0482	-363.8911	-408.6075	-2.1242	-2.2588
0.6687	2.97	183	0.6664	-0.0698	-0.1297	0.6667	0.0600	-364.0182	-408.6165	-2.1219	-2.2568

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.1+cu121
• Datasets 2.14.6
• Tokenizers 0.15.0

Training procedure

The following bitsandbytes quantization config was used during training:

• quant_method: bitsandbytes
• load_in_8bit: False
• load_in_4bit: True
• llm_int8_threshold: 6.0
• llm_int8_skip_modules: None
• llm_int8_enable_fp32_cpu_offload: False
• llm_int8_has_fp16_weight: False
• bnb_4bit_quant_type: nf4
• bnb_4bit_use_double_quant: False
• bnb_4bit_compute_dtype: float16

Framework versions

• PEFT 0.6.1