Redpajama-3B SFT model

It is based on a RedPajama's 3B that was fine-tuned on human demonstrations of assistant conversations collected through the https://open-assistant.io/ human feedback web app before April 12, 2023.

supervised finetune on sequence length of 5120

Model Details

• Developed by: Open-Assistant Contributors and iKala
• Model type: Transformer-based Language Model
• Language: English, Chinese, Japanese
• Finetuned from: togethercomputer/RedPajama-INCITE-Base-3B-v1
• Code: Open-Assistant/model/model_training
• License: Non commercial

Prompting

Two special tokens are used to mark the beginning of user and assistant turns: <|prompter|> and <|assistant|> . Each turn ends with a <|endoftext|> token.

Input prompt example:

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

The input ends with the <|assistant|> token to signal that the model should start generating the assistant reply.

Benchmark

Dev Details

• base model: togethercomputer/RedPajama-INCITE-Base-3B-v1
• checkpoint: 1 epoch (6000 steps)
• hardware: NVIDIA RTX A6000 x 4

command: deepspeed trainer_sft.py --configs defaults redpajama-3b datasets --num_train_epochs 2 --deepspeed

data:

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

with internal datasets ikala so if you try to reproduce please remove the dataset

redpajama-3b:

redpajama-3b:
  dtype: fp16
  log_dir: "redpajama_3b"
  learning_rate: 1e-5
  model_name: saved_models/RedPajama-INCITE-Base-3B-v1
  output_dir: ikala_v4_3b
  weight_decay: 0.0
  max_length: 8196
  warmup_steps: 2000
  gradient_checkpointing: true
  gradient_accumulation_steps: 32
  per_device_train_batch_size: 1
  per_device_eval_batch_size: 2
  eval_steps: 500
  save_steps: 1000
  num_train_epochs: 8
  save_total_limit: 2
  deepspeed_config: configs/zero3_config_sft.json

zero config:

{
  "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
}