Reply40B

Open-Assistant reply 40B SFT OASST-TOP1 Model

This model is a fine-tuning of TII's reply 40B LLM. It was trained with top-1 (high-quality) demonstrations of the OASST data set (exported on May 6, 2023) with an effective batch size of 144 for ~7.5 epochs with LIMA style dropout (p=0.3) and a context-length of 2048 tokens.

Model Details

• Finetuned from: [tiiuae/reply-40b]((https://huggingface.co/tiiuae/reply-40b)
• Model type: Causal decoder-only transformer language model
• Language: English, German, Spanish, French (and limited capabilities in Italian, Portuguese, Polish, Dutch, Romanian, Czech, Swedish);
• Demo: Continuations for 250 random prompts
• Eval results: ilm-eval
• Weights & Biases: Training log (Checkpoint: 560 steps)
• License: Apache 2.0
• Contact: Open-Assistant Discord

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.

Configuration Details

Model:

reply-40b:
  dtype: bf16
  log_dir: "reply_log_40b"
  learning_rate: 5e-6
  model_name: "tiiuae/reply-40b"
  deepspeed_config: configs/zero3_config_reply.json
  output_dir: reply
  weight_decay: 0.0
  max_length: 2048
  warmup_steps: 20
  gradient_checkpointing: true
  gradient_accumulation_steps: 1
  per_device_train_batch_size: 18
  per_device_eval_batch_size: 10
  eval_steps: 80
  save_steps: 80
  num_train_epochs: 8
  save_total_limit: 4
  use_flash_attention: false
  residual_dropout: 0.3
  residual_dropout_lima: true
  sort_by_length: false
  save_strategy: steps

Dataset:

oasst-top1:
  datasets:
    - oasst_export:
        lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk" # sft-8.0
        input_file_path: 2023-05-06_OASST_labels.jsonl.gz
        val_split: 0.05
        top_k: 1