Trained on the decapoda-research/llama-7b-hf base model. Does not use any inputs afaik, its all combined into the main <|user|> instruction.

Trained on 70K instructions.

Since it uses a combination of datasets, other formats seemed to have leaked in.

Some ### Instruction: & ### Response: stuff after <|user|> exists in the training. Its main training is the below format though.

Prompt:

<|user|>
What is the capital of Canada?
<|assistant|>
{The model will output here}

Training Information:

{
  "train_runtime": 20415.3794,
  "train_samples_per_second": 10.286,
  "train_steps_per_second": 10.286,
  "total_flos": 1.571869533991207e+18,
  "train_loss": 0.487219396950623,
  "epoch": 3.0,
  "step": 209994
}

{
  "micro_batch_size": 1,
  "gradient_accumulation_steps": 1,
  "num_train_epochs": 3,
  "learning_rate": 0.0003,
  "cutoff_len": 512,
  "val_set_size": 0,
  "lora_r": 16,
  "lora_alpha": 16,
  "lora_dropout": 0.05,
  "lora_target_modules": [
    "q_proj",
    "v_proj",
    "k_proj",
    "o_proj"
  ],
  "lora_modules_to_save": [],
  "train_on_inputs": true,
  "group_by_length": false,
  "load_in_8bit": false,
  "fp16": true,
  "bf16": false,
  "gradient_checkpointing": false,
  "save_steps": 10000,
  "save_total_limit": 10,
  "logging_steps": 10,
  "additional_training_arguments": null,
  "additional_lora_config": null,
}