dvilasuero's picture
dvilasuero HF staff
Update README.md
cb77620 verified
metadata
license: apache-2.0
base_model: alignment-handbook/zephyr-7b-sft-full
tags:
  - generated_from_trainer
model-index:
  - name: zephyr-7b-spin-iter0-v0
    results: []
datasets:
  - argilla/10k_prompts_SPIN_iter0_zephyr_top
  - DIBT/10k_prompts_ranked

zephyr-7b-spin-iter0-v0

This model is a fine-tuned model with SPIN starting with alignment-handbook/zephyr-7b-sft-full on the argilla/10k_prompts_SPIN_iter0_zephyr_top dataset.

It achieves the following results on the evaluation set:

  • Loss: 0.2359
  • Rewards/real: 1.3255
  • Rewards/generated: -0.8966
  • Rewards/accuracies: 0.9792
  • Rewards/margins: 2.2221
  • Logps/generated: -309.8145
  • Logps/real: -304.9670
  • Logits/generated: -2.7558
  • Logits/real: -2.7547

MT-Bench results

Model 1st Turn Score 2nd Turn Score Average Score
zephyr-7b-sft-full 6.6625 6.0250 6.34375
zephyr-7b-spin-iter0-v0 6.64375 6.1750 6.409375
zephyr-7b-spin-iter1-v0 6.90625 6.3000 6.603125
zephyr-7b-spin-iter2-v0 7.1375 6.3125 6.725000
zephyr-7b-spin-iter3-v0 7.09375 6.4500 6.771875

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

SPIN

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 5e-07
  • train_batch_size: 8
  • eval_batch_size: 8
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 4
  • gradient_accumulation_steps: 2
  • total_train_batch_size: 64
  • total_eval_batch_size: 32
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_ratio: 0.1
  • num_epochs: 2.0

Training results

Training Loss Epoch Step Validation Loss Rewards/real Rewards/generated Rewards/accuracies Rewards/margins Logps/generated Logps/real Logits/generated Logits/real
0.3011 0.96 25 0.2442 1.1606 -0.9851 0.9792 2.1457 -310.6989 -306.6157 -2.7644 -2.7641
0.0376 1.92 50 0.2359 1.3255 -0.8966 0.9792 2.2221 -309.8145 -304.9670 -2.7558 -2.7547

Framework versions

  • Transformers 4.37.0
  • Pytorch 2.1.2+cu121
  • Datasets 2.14.6
  • Tokenizers 0.15.2