afterimage 0.14.3

Generate conversational, tool-calling, structured-output, and preference datasets — easily and at scale

AfterImage

AfterImage is a Python library and CLI for generating synthetic conversational datasets with modern LLMs (Gemini, OpenAI-compatible APIs, DeepSeek, and local OpenAI-compatible servers). It is built so you can start with a YAML file and one command, then compose callbacks, document providers, storage, evaluation, and export pipelines as your needs grow—from quick experiments to large, production-style runs.

Two ways to work (same engine)

1. CLI and config — easy to begin
Describe generation in YAML, set your API key in the environment, and run afterimage generate. No boilerplate, no custom harness required to get JSONL on disk. Optional commands cover export to fine-tuning formats and preference (DPO-style) pair generation.

2. Python API — composable and extensible
Use ConversationGenerator, StructuredGenerator, and PersonaGenerator with pluggable instruction generators, respondent prompt modifiers, stopping criteria, storage (JSONL or SQL), quality judges, and monitoring. The same abstractions power the CLI; you swap or combine pieces instead of forking the stack.

That split keeps onboarding shallow while leaving room for scale (concurrency, key pools, SQL storage) and specialized flows (RAG-style context, personas, structured extraction, preference data). Guides and API reference are on afterimage.altai.dev.

---

Installation

The package can be installed from PyPI as afterimage.

uv add afterimage

pip install afterimage

Optional extras (see pyproject.toml for exact dependency sets):

uv add "afterimage[embeddings-local]"
# or
pip install "afterimage[embeddings-local]"

Extra	Purpose
embeddings-local	Local embeddings (sentence-transformers) for process-based embedding providers, Qdrant-style workflows, and quality checks that need a local model.
server	FastAPI app (afterimage-server entry point).
training	Torch / TRL stack, Gradio, and FastMCP for examples/demo_ui and the training scripts under examples/.

---

Start in minutes (CLI)

Requires Python 3.11+ and an API key (e.g. GEMINI_API_KEY for the sample config).

afterimage generate -c examples/configs/basic.yaml

Dry-run the plan without calling the API:

afterimage generate -c examples/configs/basic.yaml --dry-run

Export a dataset to common fine-tuning formats:

afterimage export -i your_dataset.jsonl -f sharegpt -f messages
afterimage export --list-formats

Generate preference pairs from config:

afterimage preference -c your_config.yaml

More examples live under examples/configs/. In-depth guides (conversations, personas, structured generation, evaluation, export, preference data, local models) are on afterimage.altai.dev.

---

What you can build

• Multi-turn synthetic chat for SFT, evaluation sets, or simulation.
• Document-grounded questions and answers (instruction side + optional respondent context).
• Persona-driven diversity tied to your corpus.
• Structured outputs via Pydantic schemas (single-turn extraction or generation).
• DPO / RLHF-style preference data with multiple variation strategies.
• Quality loops (async judge, optional auto-improve) and observability (metrics, periodic alert checks, exports to JSON/CSV/Parquet).

---

Repository layout

Path	Contents
docs/	Sphinx sources; mirrors and extends the hosted site.
examples/	YAML configs and demo flows.
DESIGN.md	Architecture and design notes for contributors.
afterimage/	Library and CLI implementation.

Source & issues: github.com/altaidevorg/afterimage

---

License

Apache License 2.0 (see the LICENSE file and PyPI package metadata).