README.md

Haha

Haha is an LLM finetune for generating punchlines to joke setups, as an experiment in creative writing.

Training Pipeline

The Synagi training pipeline uses Axolotl and automatically deploys HuggingFace models with corresponding Weights & Biases data, using the configs and datasets merged to the main branch of this repo.

Links

• Model YAML Configs: Each config represent a set of hyperparameters for the Haha fintune
• Finetune Datasets: Data used for finetuning in Alpaca format
• HuggingFace models: Trained models, using ChatML prompt format
• Weights & Biases data: Training results
• Eval configs: Configs used for evaluating the trained models
• Results: Per-config results for evals on each model

Configs

Each config file in /train/config represents a planned unique finetune of the base model, as a Lora or QLora.

This enables configs to be committed here, that then get trained and uploaded and with all their training results available.

The config's basename (name without file extension) is shared between its config here, its WandB run name, and a corresponding HuggingFace model upload.

Config Format

• The configs need to be named consecutively with 3 digits at the end, e.g. haha-llama-3-8b-lora-001.yml.
• The WandB name for the training run is set dynamically within the pipeline to the config file name, so leave it blank in input the config. The project name should be 'haha'.
• A Lora adapter of the training is uploaded to HuggingFace, as a model with the same basename as the config.

Example

base_model: meta-llama/Meta-Llama-3-8B
...

chat_template: chatml
default_system_message: You are a creative and hilarious comedy writer that loves to craft jokes.

datasets:
  - path: /workspace/data/haha-v1.jsonl
    type:
      system_prompt: "system\nYou are a creative and hilarious comedy writer that loves to craft jokes.<|im_end|>\n"
      field_system: system
      field_instruction: instruction
      field_output: output
      format: "<|im_start|>user\n{instruction}<|im_end|>\n<|im_start|>assistant\n"
      no_input_format: "<|im_start|>user\n{instruction}<|im_end|>\n<|im_start|>assistant\n"

dataset_prepared_path: ./last_run_prepared
val_set_size: 0.01
output_dir: /workspace/output
...

Datasets

Datasets should be stored in the /data folder, and then referenced within the configs in a /workspace/data/ folder. The pipeline will use them for the training.

Config Format

The dataset formats should be Alpaca, and are trained using the ChatML prompt format.

Example

{"instruction": "In a survey of 35 cities, Los Angeles ranked second to last in \"intelligence.\"", "output": "Residents of LA were outraged after the report was slowly explained to them.", "input": ""}

Eval

Edit the /eval/eval.json to define eval settings using the following sections:

1. prompts: An array of strings, the joke setup prompts to test.

2. settings: An array of objects, parameters for the inference, such as temperature (which influences the randomness of the model's responses).

3. seedVariants: An integer, number of times prompts for a specific setting should be repeated with different seed numbers, starting at 1.

Example

{
    "prompts": [
        "It was 100 degrees today.",
        "An iPad used by the Pope is being auctioned off for charity.",
        ...
    ],
    "settings": [
        { "temperature": 0.1 },
        { "temperature": 0.3 },
        ...
    ],
    "seedVariants": 3
}

Results

Results for each config eval are stored as individual JSON files in the /results folder, with the following structure:

Note: "default_system_message" value in the config YAML is used as the system prompt in that model's eval.

1. config: The config name and its key settings used in the eval, including system prompt.
2. settingsResults: An array of objects for each set of model settings tested against the prompt. Each object includes:
   • temperature: Temperature setting used during inference
   • completions: An array of the model's completions, in order of seed number starting at 1

Example

{
  "config": {
    "name": "haha-llama-3-8b-lora-001",
    "default_system_message": "You are a creative and hilarious comedy writer that loves to craft jokes.",
    "sequence_len": 64,
    "num_epochs": 4,
    "optimizer": "adamw_bnb_8bit",
    "lr_scheduler": "cosine",
    "learning_rate": 0.00005
  },
  "results": [
    {
      "prompt": "It was 100 degrees today.",
      "settingsResults": [
        {
          "temperature": 0.1,
          "completions": [
            "Response for seed 1",
            "Response for seed 2",
            "Response for seed 3"
          ]
        },
        ...
      ]
    },
    ...
  ]
}