Bloom-3B SFT模型

它是基于Bloom-zh的3B模型，在2023年4月12日之前通过人类反馈网络应用收集的助理对话的人类示范进行微调的。

监督微调的序列长度为5120

模型详情

• 开发者： Open-Assistant Contributors 和 iKala
• 模型类型：基于Transformer的语言模型
• 语言：英语、中文、日语
• 微调自： ckip-joint/bloom-3b-zh
• 代码： Open-Assistant/model/model_training
• 许可证： MEDIATEK RESEARCH许可证（ link ）和RAIL许可证v1.0（ link ），非商业用途

提示

有两个特殊的令牌用于标记用户和助手的开始回合：<|prompter|>和<|assistant|>。每个回合以</s>令牌结束。

输入示例提示：

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

输入以<|assistant|>令牌结束，表示模型应开始生成助手回复。

基准

开发详情

• 基本模型： ckip-joint/bloom-3b-zh
• 检查点：1个时期（6000个步骤）
• 硬件：NVIDIA RTX A6000 x 4

命令：deepspeed trainer_sft.py --configs defaults bloom-zh-3b datasets --num_train_epochs 2 --deepspeed

数据：

datasets:
  - wmt2019_zh-en:
      max_val_set: 1000
      max_train_set: 20000
  - ted_trans_en-ja:
      max_val_set: 1000
      max_train_set: 20000
  - ted_trans_zh-ja:
      max_val_set: 1000
      max_train_set: 20000
  - ikala:
      input_file_path: export_conversation_v4.4.jsonl
      val_split: 0.05
  - dolly15k:
      val_split: 0.05
  - oasst_export:
      lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk,zh,ja,th,ko"
      input_file_path: 2023-04-12_oasst_release_ready_synth.jsonl.gz
      val_split: 0.05
  - joke
  - gsm8k
  - webgpt

使用内部数据集ikala，如果要复现，请删除数据集

bloom-zh-3b：

bloom-zh-3b:
  dtype: fp16
  log_dir: "bloom-zh_3b"
  learning_rate: 8e-6
  model_name: ckip-joint/bloom-3b-zh
  output_dir: bloom_model_v4_3b
  weight_decay: 0.0
  max_length: 5120
  warmup_steps: 2000
  gradient_checkpointing: true
  gradient_accumulation_steps: 32
  per_device_train_batch_size: 1
  per_device_eval_batch_size: 1
  eval_steps: 500
  save_steps: 1000
  num_train_epochs: 8
  save_total_limit: 2
  deepspeed_config: configs/zero3_config_sft.json

零配置：

{
  "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",
      "warmup_type": "linear",
      "total_num_steps": "auto"
    }
  },
  "zero_optimization": {
    "stage": 3,
    "overlap_comm": true,
    "contiguous_gradients": true,
    "sub_group_size": 1e9,
    "reduce_bucket_size": "auto",
    "stage3_prefetch_bucket_size": "auto",
    "stage3_param_persistence_threshold": "auto",
    "stage3_max_live_parameters": 1e9,
    "stage3_max_reuse_distance": 1e9,
    "stage3_gather_16bit_weights_on_model_save": 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
}