Skip to content

delphi-suite/delphi

Repository files navigation

delphi

delphi is a set of tools for standardized and (mostly) reproducible training of small language models. You can use delphi to train a custom tokenizer, tokenize your dataset, and train your model. We build on top of HuggingFace, supporting every CausalLM architecture. Datasets, tokenizers and models (including checkpoints!) can be downloaded from and uploaded to HuggingFace automatically, with no need to manage local files.

Setup

  1. Clone the repo
git clone https://github.com/delphi-suite/delphi.git
cd delphi  
  1. Make & activate python >= 3.10 virtual env
python3.10 -m venv .venv
source .venv/bin/activate
  1. Install the project in editable state
    pip install -e .
    See [project.optional-dependencies] section in pyproject.toml for additional dependencies, e.g. you may want to pip install -e ."[dev,mamba_cuda]"
  2. get your HuggingFace and W&B tokens and put them in the environment variables
export HF_TOKEN=...
export WANDB_API_KEY=...

Training a tokenizer

If you want to train a small and efficient model on a narrow dataset, then we recommend using a custom tokenizer with a small vocabulary. To train a reversible, GPT2-style, BPE tokenizer you can use scripts/train_tokenizer.py.

Script usage:

> scripts/train_tokenizer.py --help
usage: train_tokenizer.py [-h] --in-dataset IN_DATASET --feature FEATURE --split SPLIT
                          --vocab-size VOCAB_SIZE
                          [--out-dir OUT_DIR] [--out-repo OUT_REPO]

Train a custom, reversible, BPE tokenizer (GPT2-like). You need to provide --out-repo or --out-dir.

options:
  -h, --help            show this help message and exit
  --in-dataset IN_DATASET, -i IN_DATASET
                        Dataset you want to train the tokenizer on. Local path or HF repo id
  --feature FEATURE, -f FEATURE
                        Name of the feature (column) containing text documents in the input dataset
  --split SPLIT, -s SPLIT
                        Split of the dataset to be used for tokenizer training, supports slicing like 'train[:10%]'
  --vocab-size VOCAB_SIZE, -v VOCAB_SIZE
                        Vocabulary size of the tokenizer
  --out-dir OUT_DIR     Local directory to save the resulting tokenizer
  --out-repo OUT_REPO   HF repo id to upload the resulting tokenizer

Here's how we trained the tokenizer for our stories-* suite of models. Please note that you can use single letter abbreviations for most arguments.

> scripts/train_tokenizer.py \
    --in-dataset delphi-suite/stories \
    --feature story \
    --split train \
    --vocab-size 4096 \
    --out-repo delphi-suite/stories-tokenizer

We use the only feature named story in the train split of delphi-suite/stories. We train a tokenizer with a vocabulary of 4096 tokens, and upload it to HF model repo delphi-suite/stories-tokenizer.

Tokenizing a dataset

To turn a collection of text documents into sequences of tokens required for model training, you can use scripts/tokenize_dataset.py. All documents are tokenized and concatenated, with the <eos> token as a separator, e.g.

doc1_tok1, doc1_tok2, ..., doc1_tokX, <eos>, doc2_tok1, doc2_tok2, ..., doc2_tokX, <eos>, doc3_tok1, ...

Then this is divided into chunks, and the <bos> token is inserted at the begining of each chunk, e.g.

<bos> doc1_tok1, doc1_tok2, ..., doc1_tokX, <eos>, doc2_tok1
<bos> doc2_tok2, ..., doc2_tok511
<bos> doc2_tok512, doc2_tok513, ..., doc2_tokX <eos>, doc3_tok1, ...
...

It will produce sequences of specified size, by discarding the last chunk if it's too short. We don't use padding.

Script usage:

> scripts/tokenize_dataset.py --help
usage: tokenize_dataset.py [-h] --in-dataset IN_DATASET --feature FEATURE --split SPLIT
                           --tokenizer TOKENIZER --seq-len SEQ_LEN
                           [--batch-size BATCH_SIZE] [--chunk-size CHUNK_SIZE]
                           [--out-dir OUT_DIR] [--out-repo OUT_REPO]

Tokenize a text dataset using a specific tokenizer

options:
  -h, --help            show this help message and exit
  --in-dataset IN_DATASET, -i IN_DATASET
                        Dataset you want to tokenize. Local path or HF repo id
  --feature FEATURE, -f FEATURE
                        Name of the feature (column) containing text documents in the input dataset
  --split SPLIT, -s SPLIT
                        Split of the dataset to be tokenized, supports slicing like 'train[:10%]'
  --tokenizer TOKENIZER, -t TOKENIZER
                        HF repo id or local directory containing the tokenizer
  --seq-len SEQ_LEN, -l SEQ_LEN
                        Length of the tokenized sequences
  --batch-size BATCH_SIZE, -b BATCH_SIZE
                        How many text documents to tokenize at once (default: 50)
  --chunk-size CHUNK_SIZE, -c CHUNK_SIZE
                        Maximum number of tokenized sequences in a single parquet file (default: 200_000)
  --out-dir OUT_DIR     Local directory to save the resulting dataset
  --out-repo OUT_REPO   HF repo id to upload the resulting dataset

Here's how we tokenized the dataset for our stories-* suite of models. Please note that you can use single letter abbreviations for most arguments.

For train split:

> scripts/tokenize_dataset.py \
    --in-dataset delphi-suite/stories \
    --feature story \
    --split train \
    --tokenizer delphi-suite/stories-tokenizer \
    --seq-len 512 \
    --out-repo delphi-suite/stories-tokenized

For validation split, repeated arguments omitted:

> scripts/tokenize_dataset.py \
    ...
    --split validation \
    ...

The input dataset is the same as in tokenizer training example above. We tokenize it with our custom delphi-suite/stories-tokenizer into sequences of length 512. We upload it to HF dataset repo delphi-suite/stories-tokenized.

Please note that you can use any HuggingFace tokenizer, you don't need to train a custom one.

Training a model

To train a model, you'll need to create a config file. For examples see configs/, and for field descriptions see delphi/train/config/training_config.py. The training script is located in scripts/train_model.py.

Script usage:

> scripts/train_model.py --help
usage: train_model.py [-h] [--overrides [OVERRIDES ...]] [-v | -s] [config_files ...]

Train a delphi model

positional arguments:
  config_files          Path to json file(s) containing config values, e.g. 'primary_config.json secondary_config.json'.

options:
  -h, --help            show this help message and exit
  --overrides [OVERRIDES ...]
                        Override config values with space-separated declarations. e.g. `--overrides model_config.hidden_size=42 run_name=foo`
  -v, --verbose         Increase verbosity level, repeatable (e.g. -vvv). Mutually exclusive with --silent, --loglevel
  -s, --silent          Silence all logging. Mutually exclusive with --verbose, --loglevel

You can specify primary config and secondary config, which is useful if you're training a suite of models that only differ in a few parameters. Additionally, you can override specific fields using the --overrides flag. If you don't want to push the model and its checkpoints to HF, you need to explicitly set out_repo="". If you don't want to log to W&B, you need to set wandb="". Please note that by default we save the optimizer state (2x model size) with every checkpoint.

Here is how we trained our stories-mamba-100k model

> scripts/train_model.py \
    configs/stories/mamba/base.json \
    configs/stories/mamba/100k.json \
    --overrides \
      out_repo="delphi-suite/stories-mamba-100k" \
      wandb="delphi-suite/delphi"

Development

  1. Install the dev and notebooks dependencies pip install -e ."[dev,notebooks]".
  2. Run the tests pytest.
  3. Install pre-commit pre-commit install.
  4. Install the recommended vscode extensions.

When you save a file vscode should automatically format it. Otherwise, pre-commit will do that, but you will need to add the changes and commit again.

Citation

If you use delphi in your research, please cite using the following

@misc{delphi,
  title = {delphi: small language models training made easy},
  author = {Janiak, J. and Dhyani, J. and Brinkmann, J. and Paulo, G. and Wendland, J. and Alonso, V. A. and Li, S. and Duong, P. A. and Rigg, A.},
  year = 2024,
  url = {https://github.com/delphi-suite/delphi},
}