Oasst Sft 4 Pythia 12b Epoch 3.5

🚀 Open-Assistant SFT-4 12B模型

这是Open-Assistant项目的第4代英文监督微调（SFT）模型。它基于Pythia 12B，在2023年3月25日之前通过https://open-assistant.io/人工反馈网页应用收集的人工对话示范数据上进行了微调。

✨ 主要特性

• 基于Transformer架构的语言模型，能够处理文本生成任务。
• 经过人工对话示范数据的监督微调，具备更好的对话生成能力。

📦 安装指南

文档未提供安装步骤，故跳过此章节。

💻 使用示例

基础用法

使用两个特殊标记 <|prompter|> 和 <|assistant|> 分别标记用户和助手的对话起始，每个对话轮次以 <|endoftext|> 标记结束。

输入提示示例：

<|prompter|>What is a meme, and what's the history behind this word?<|endoftext|><|assistant|>

输入以 <|assistant|> 标记结尾，用于提示模型开始生成助手回复。

📚 详细文档

模型详情

开发详情

• wandb: https://wandb.ai/open-assistant/supervised-finetuning/runs/770a0t41
• 基础模型: andreaskoepf/pythia-12b-pre-2000
• 检查点: 4000步

命令: deepspeed trainer_sft.py --configs defaults reference-data reference-pythia-12b --cache_dir /home/ubuntu/data_cache --output_dir .saved/oasst-sft-3-pythia-12b-reference_2kpre --num_train_epochs 8 --residual_dropout 0.2 --deepspeed --use_flash_attention true --model_name andreaskoepf/pythia-12b-pre-2000

数据:

reference-data:
  datasets:
    - oasst_export:
      lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk"
      input_file_path: 2023-03-25_oasst_research_ready_synth_labels.jsonl.gz
      val_split: 0.05
    - alpaca
  sort_by_length: false
  use_custom_sampler: false

pythia配置:

reference-pythia-12b:
  dtype: fp16
  log_dir: "pythia_log_12b"
  learning_rate: 6e-6
  model_name: EleutherAI/pythia-12b-deduped
  output_dir: pythia_model_12b
  weight_decay: 0.0
  max_length: 2048
  warmup_steps: 100
  gradient_checkpointing: true
  gradient_accumulation_steps: 2
  per_device_train_batch_size: 4
  per_device_eval_batch_size: 4
  eval_steps: 100
  save_steps: 1000
  num_train_epochs: 8
  save_total_limit: 4

zero配置:

{
  "fp16": {
    "enabled": "auto",
    "loss_scale": 0,
    "loss_scale_window": 1000,
    "initial_scale_power": 16,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "bf16": {
    "enabled": "auto"
  },
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": "auto",
      "betas": "auto",
      "eps": "auto",
      "weight_decay": "auto"
    }
  },
  "scheduler": {
    "type": "WarmupDecayLR",
    "params": {
      "warmup_min_lr": "auto",
      "warmup_max_lr": "auto",
      "warmup_num_steps": "auto",
      "total_num_steps": "auto"
    }
  },
  "zero_optimization": {
    "stage": 2,
    "allgather_partitions": true,
    "allgather_bucket_size": 1e9,
    "overlap_comm": false,
    "reduce_scatter": true,
    "reduce_bucket_size": 1e9,
    "contiguous_gradients": true
  },
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "steps_per_print": 2000,
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "wall_clock_breakdown": false
}

📄 许可证

本模型使用Apache 2.0许可证。