metadata

license: gemma
language:
  - en
pipeline_tag: text-generation
tags:
  - gemma
  - gguf
  - SPPO
  - imatrix
base_model: UCLA-AGI/Gemma-2-9B-It-SPPO-Iter3

Quant Infos

Updated for all recent llama.cpp fixes (final logit soft capping+sliding window+tokenizer)

quants done with an importance matrix for improved quantization loss
Requantized ggufs & imatrix from hf bf16
- initial version was based on f32 gguf provided by google, which had various issues
- also updated for all recent llama.cpp fixes (final logit soft capping+sliding window+tokenizer)
Wide coverage of different gguf quant types from Q_8_0 down to IQ1_S
- experimental custom quant types
  - _L with --output-tensor-type f16 --token-embedding-type f16 (same as bartowski's)
Quantized with llama.cpp commit 5fac350b9cc49d0446fc291b9c4ad53666c77591 (master from 2024-07-02)

Imatrix generated with this multi-purpose dataset by bartowski.

./imatrix -m $model_name-bf16.gguf -f calibration_datav3.txt -o $model_name.imatrix

Original Model Card:

Self-Play Preference Optimization for Language Model Alignment (https://arxiv.org/abs/2405.00675)

Gemma-2-9B-It-SPPO-Iter3

This model was developed using Self-Play Preference Optimization at iteration 3, based on the google/gemma-2-9b-it architecture as starting point. We utilized the prompt sets from the openbmb/UltraFeedback dataset, splited to 3 parts for 3 iterations by snorkelai/Snorkel-Mistral-PairRM-DPO-Dataset. All responses used are synthetic.

Terms of Use: Terms

Links to Other Models

Model Description

Model type: A 8B parameter GPT-like model fine-tuned on synthetic datasets.
Language(s) (NLP): Primarily English
License: Apache-2.0
Finetuned from model: google/gemma-2-9b-it

AlpacaEval Leaderboard Evaluation Results

Model	LC. Win Rate	Win Rate	Avg. Length
Gemma-2-9B-SPPO Iter1	48.70	40.76	1669
Gemma-2-9B-SPPO Iter2	50.93	44.64	1759
Gemma-2-9B-SPPO Iter3	53.27	47.74	1803

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 5e-07
eta: 1000
per_device_train_batch_size: 8
gradient_accumulation_steps: 1
seed: 42
distributed_type: deepspeed_zero3
num_devices: 8
optimizer: RMSProp
lr_scheduler_type: linear
lr_scheduler_warmup_ratio: 0.1
num_train_epochs: 1.0

Citation

@misc{wu2024self,
      title={Self-Play Preference Optimization for Language Model Alignment}, 
      author={Wu, Yue and Sun, Zhiqing and Yuan, Huizhuo and Ji, Kaixuan and Yang, Yiming and Gu, Quanquan},
      year={2024},
      eprint={2405.00675},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}