Edit model card

stealth-finance-v2-dpo-adapter

This model is a fine-tuned version of TomGrc/FusionNet_7Bx2_MoE_v0.1 on the jan-hq/distilabel_dpo_pairs_binarized, the argilla/OpenHermes2.5-dpo-binarized-alpha, the jan-hq/capybara_dpo_binarized, the jan-hq/bagel_dpo_binarized, the jan-hq/ultrafeedback_preferences_cleaned_binarized, the jan-hq/openmath_instruct_dpo_binarized, the jan-hq/distil_math_dpo_binarized and the jan-hq/evol_codealpaca_dpo_binarized datasets. It achieves the following results on the evaluation set:

  • Loss: 0.1290
  • Rewards/chosen: -0.1799
  • Rewards/rejected: -6.0696
  • Rewards/accuracies: 0.8597
  • Rewards/margins: 5.8897
  • Logps/rejected: -324.0384
  • Logps/chosen: -275.3572
  • Logits/rejected: -0.7749
  • Logits/chosen: -0.7773

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-05
  • train_batch_size: 1
  • eval_batch_size: 1
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 2
  • gradient_accumulation_steps: 32
  • total_train_batch_size: 64
  • total_eval_batch_size: 2
  • 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

Training results

Training Loss Epoch Step Validation Loss Rewards/chosen Rewards/rejected Rewards/accuracies Rewards/margins Logps/rejected Logps/chosen Logits/rejected Logits/chosen
0.3593 1.0 3280 0.1290 -0.1799 -6.0696 0.8597 5.8897 -324.0384 -275.3572 -0.7749 -0.7773

Framework versions

  • PEFT 0.8.2
  • Transformers 4.37.2
  • Pytorch 2.1.2+cu121
  • Datasets 2.14.6
  • Tokenizers 0.15.0
Downloads last month
0
Inference API
Unable to determine this model’s pipeline type. Check the docs .

Model tree for jan-hq/stealth-finance-v2-dpo-adapter

Adapter
(1)
this model

Datasets used to train jan-hq/stealth-finance-v2-dpo-adapter