GitHub 5 Install

Integrations

Parallelogram runs wherever your data does: inside your coding agent, in CI, and right before your trainer.

Coding Agents

Parallelogram runs as an MCP server, so your agent checks and fixes data itself.

Claude Code

Writes .mcp.json for this project.
shell
```
parallelogram integrate claude-code
```
Or: claude mcp add parallelogram -- parallelogram mcp
Codex

Codex keeps MCP servers in your user config; this adds it there.
shell
```
codex mcp add parallelogram \
  -- parallelogram mcp
```
parallelogram integrate codex prints the TOML instead.
Cursor

Writes .cursor/mcp.json for this project.
shell
```
parallelogram integrate cursor
```
VS Code + Copilot

Agent mode. Writes .vscode/mcp.json.
shell
```
parallelogram integrate vscode
```
Gemini CLI

Writes .gemini/settings.json for this project.
shell
```
parallelogram integrate gemini
```
Windsurf

Prints the entry for Windsurf's MCP config.
shell
```
parallelogram integrate windsurf
```
Cline

Prints the entry for Cline's MCP settings.
shell
```
parallelogram integrate cline
```
Zed

Prints the context server for Zed's settings.
shell
```
parallelogram integrate zed
```
Any MCP client

A stdio server: command parallelogram, args ["mcp"].
shell
```
parallelogram mcp
```
Tools: check_dataset, fix_dataset, list_rules.

CI and Hooks

Bad data fails the build, the same way broken code does.

GitHub Actions

Fails a pull request on errors and writes a report to the job summary.
yaml
```
- uses: Thatayotlhe04/Parallelogram@v1
  with:
    path: data/train.jsonl
```

pre-commit

Checks every staged .jsonl before it's committed.

yaml

- repo: https://github.com/Thatayotlhe04/Parallelogram
  rev: cli-v1.6.0
  hooks:
    - id: parallelogram

Any CI

Exit 0 clean, 1 warnings, 2 errors, 3 operational failure.
shell
```
parallelogram check data/train.jsonl \
  --json --limit 50
```

Training Frameworks

The formats each trainer reads, checked before the GPU clock starts.

Hugging Face TRL

SFT messages and DPO prompt/chosen/rejected. Exact counts with the model's tokenizer.
shell
```
parallelogram check train.jsonl \
  --tokenizer Qwen/Qwen2.5-7B
```
Axolotl

Chat-template, ShareGPT and Alpaca datasets. Catches records over sequence_len.
shell
```
parallelogram check train.jsonl \
  --max-seq-len 4096
```
Unsloth

ShareGPT and chat data, checked against your max_seq_length before the SFT run starts.
shell
```
parallelogram check train.jsonl \
  --max-seq-len 2048
```
LLaMA-Factory

Alpaca and ShareGPT, including function_call and observation turns.
shell
```
parallelogram check data/train.jsonl
```
OpenAI chat format

messages with tools and tool_calls: the shape most trainers accept, OpenAI's fine-tuning API included.
shell
```
parallelogram check train.jsonl \
  --format openai-chat
```

Python

The same checks inside your own pipeline.

python

from parallelogram import check

result = check("train.jsonl")
if result.exit_code >= 2:
    raise SystemExit(result.exit_code)