README.md

metadata

pretty_name: Go Fix SFT
language:
  - code
license:
  - apache-2.0
  - mit
  - bsd-3-clause
  - bsd-2-clause
  - isc
task_categories:
  - text-generation
tags:
  - go
  - source-code
  - supervised-fine-tuning
configs:
  - config_name: chatml
    data_files:
      - split: train
        path: chatml/train-*
      - split: validation
        path: chatml/validation-*
      - split: test
        path: chatml/test-*
  - config_name: default
    data_files:
      - split: train
        path: examples/split=train/*.parquet
      - split: validation
        path: examples/split=validation/*.parquet
      - split: test
        path: examples/split=test/*.parquet
dataset_info:
  config_name: chatml
  features:
    - name: pair_id
      dtype: string
    - name: repository
      dtype: string
    - name: commit
      dtype: string
    - name: path
      dtype: string
    - name: license
      dtype: string
    - name: original_content_sha256
      dtype: string
    - name: fixed_content_sha256
      dtype: string
    - name: near_cluster_id
      dtype: string
    - name: messages
      list:
        - name: content
          dtype: string
        - name: role
          dtype: string
  splits:
    - name: train
      num_bytes: 1401554066
      num_examples: 78357
    - name: validation
      num_bytes: 10824338
      num_examples: 591
    - name: test
      num_bytes: 10994187
      num_examples: 540
  download_size: 372210965
  dataset_size: 1423372591

Go Fix SFT

Go Fix SFT is a supervised fine-tuning dataset of validated original-to-modernized Go source pairs. Each output was produced with Go 1.26 go fix, is gofmt canonical, and passed a post-transformation package check.

Dataset size

Split	Examples
Train	78,357
Validation	591
Test	540
Total	79,488

The examples come from 2,501 repositories. Splits are assigned by repository and checked for near-clone leakage.

Fields

The main training fields are:

• system_prompt: stable model behavior instruction
• instruction: requested modernization task
• input_content: complete original Go source file
• output_content: complete modernized Go source file

Rows also include repository, commit, path, license, module, toolchain, content hashes, validation results, and clone-cluster provenance.

Loading

from datasets import load_dataset

dataset = load_dataset("sumpterc/go-fix-sft")

Apply the target model's chat template to system_prompt, instruction plus input_content, and output_content. Perform model-specific token-length validation before training.

Filtering policy

Only final split-eligible, non-generated, non-sensitive, gofmt-canonical transformations are included. Examples whose combined input and output exceed 65,536 bytes are not truncated; their metadata is retained in excluded.parquet.

Licenses

Each row retains the detected license and exact source repository/commit provenance. Included licenses are Apache-2.0, MIT, BSD-2-Clause, BSD-3-Clause, and ISC. The underlying source remains governed by its respective license.

Exact-commit license bodies and discovered Apache NOTICE files for all 2,501 represented repositories are included under third_party/. The human-readable consolidated text is third_party/THIRD_PARTY_LICENSES.txt; Parquet indexes retain raw bytes, paths, hashes, validation methods, and repository mappings.

Limitations

This dataset teaches the mechanical transformations validated through go fix; it is not a general-purpose Go instruction dataset. The examples are machine-generated and machine-validated, so representative human review is recommended before a production model release.

See manifest.json for complete construction policy, source hashes, counts, and artifact hashes.