Hugging Face's logo

Datasets:
archit11
/
hyperswitch-token-aware-cpt2

Modalities:
Text
Formats:
parquet
Size:
1K - 10K
Libraries:
Datasets
pandas
Polars
Dataset card Data Studio Files
xet
Community
hyperswitch-token-aware-cpt2 / README.md
archit11's picture
archit11
Upload README.md with huggingface_hub
731bfdc verified
preview code
|
raw
history blame contribute delete
2.81 kB

Hyperswitch Token-Aware CPT Dataset

This dataset contains 1,076 samples of Rust code from the Hyperswitch payment router project, optimized for Continued Pre-Training (CPT) with the Kwaipilot/KAT-Dev tokenizer.

Dataset Statistics

  • Total Samples: 1,076
  • Total Tokens: 5,687,255
  • Mean Tokens per Sample: 5,285
  • Token Range: 2,001 - 15,609

Token Distribution

  • < 4k tokens: 38.1% of samples
  • 4k-10k tokens: 52.0% of samples
  • 10k+ tokens: 9.9% of samples

Granularity Types

  • file: 721 samples (single large files)
  • module: 180 samples (multiple files from same module)
  • combined_files: 160 samples (small files combined by crate)
  • crate: 15 samples (entire small crates)

Top Crates

  1. router - 371 samples
  2. hyperswitch_connectors - 336 samples
  3. analytics - 54 samples
  4. diesel_models - 39 samples
  5. api_models - 28 samples

Sample Structure

Each sample contains:

  • id: Unique identifier
  • type: Sample type (always "clm" for causal language modeling)
  • granularity: Level of code organization (file/module/combined_files/crate)
  • content: Full code with path metadata in format:
    <path>
Repository: hyperswitch
Crate: [crate_name]
File: [file_path]
Tokens: [token_count]
</path>

<file>
[actual code content]
</file>
  • metadata: Contains crate, file info, and token count

Usage 

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("archit11/hyperswitch-token-aware-cpt")

# Access samples
sample = dataset['train'][0]
print(f"Tokens: {sample['metadata']['token_count']:,}")
print(f"Crate: {sample['metadata']['crate']}")
print(f"Granularity: {sample['granularity']}")

# Filter by token count
medium_samples = dataset['train'].filter(
    lambda x: 4000 <= x['metadata']['token_count'] < 10000
)

# Filter by crate
router_samples = dataset['train'].filter(
    lambda x: x['metadata']['crate'] == 'router'
)

Training Recommendations

  • Context Length: 16k tokens (max sample is 15,609 tokens)
  • Tokenizer: Kwaipilot/KAT-Dev
  • Suggested Batch Size: 1-2 samples per batch (due to large context)
  • Format: Samples are pre-formatted with <path> and <file> tags

Source

Generation Method

Samples were generated using token-aware strategies:

  1. Large files (2k-16k tokens) included as-is
  2. Small files combined within same crate until reaching 2k+ tokens
  3. Module clusters grouped by directory structure
  4. Complete crates for small crates that fit within context

All token counts measured using the Kwaipilot/KAT-Dev tokenizer.