postsynth 0.1.0

Generate synthetic post-training datasets in TRL-native formats.

postsynth

Generate synthetic post-training datasets in TRL-native formats.

postsynth v1 focuses on conversational JSONL rows for common TRL workflows:

• SFT: messages
• DPO: prompt, chosen, rejected
• GRPO: prompt
• KTO: prompt, completion, label

Install

uv sync --dev

Configure OpenRouter

Set your key in the environment:

export OPENROUTER_API_KEY="..."

For local development, you can also create a .env file:

OPENROUTER_API_KEY=...

.env files are ignored by git.

CLI

Generate from a seed instruction:

postsynth generate sft \
  --seed "Customer-support chats about subscription billing." \
  --count 100 \
  --out data/generated/sft.jsonl

Generate from examples:

postsynth generate dpo \
  --examples examples/dpo.jsonl \
  --count 100 \
  --out data/generated/dpo.jsonl

The CLI writes a JSONL dataset and a sibling dataset card, for example sft.dataset.md, with non-secret generation metadata.

Python API

from postsynth import generate_sft

result = generate_sft(
    seed="Math tutoring conversations for middle-school students.",
    count=25,
    output_path="data/generated/math_sft.jsonl",
)

print(result.rows)

TRL Loading

from datasets import load_dataset

dataset = load_dataset("json", data_files="data/generated/sft.jsonl")

The generated rows follow the conversational formats documented by TRL: https://huggingface.co/docs/trl/en/dataset_formats