# AfterImage — synthetic training data — ALTAI

> AfterImage produces training examples: define the task, shape the training data. Open source under Apache-2.0.

## AfterImage. Define the task. Shape the training data.

AfterImage produces training examples. Describe the task, the people who will ask and, if you have them, the source documents. It writes varied examples to review and evaluate before training.

[GitHub](https://github.com/altaidevorg/afterimage)

1. **One question:** A customer asks about the filter. The manual has the answer.
2. **Many voices:** Other personas ask it in their own words. The answer stays.
3. **One answer:** Every phrasing is paired with the one answer from the manual.

AfterImage rewrites one question in the voices of different people and pairs each with the same sourced answer. CUSTOMER: “How often should I check the filter?” NEW USER: “Never looked at the filter. When should I?” TECHNICIAN: “What is the filter inspection interval?” BUSY MANAGER: “Filter: how often?” ANGRY CUSTOMER: “Clogged again! When should I have cleaned it?” TYPOS: “how ofen chek filtr” Every one is answered: “Per the manual, check the filter every month.” The examples are written for this scene.

### Generate training data around the work a model needs to do.

AfterImage creates conversational, tool-calling, structured-output, and preference datasets from task instructions and optional source documents. It provides the data-generation layer for teaching a model a specific task, with generated examples reviewed and evaluated before training.

#### Shape examples to a task

Generate multi-turn conversations, document-grounded questions, structured outputs, and tool-call examples.

#### Create preference data

Produce preference pairs for training methods such as DPO and use the quality checks available in the generation pipeline.

#### Export for training

Review datasets and export them in formats used by common model-training workflows.

### From input to output.

1. **Input:** A task definition, optional reference documents, and the response or tool-use behavior a model should learn.
2. **Process:** Generate varied examples, inspect quality signals, and export the dataset to a separate training and evaluation workflow.
3. **Output:** A reviewable dataset of conversational, structured-output, or other synthetic examples for model adaptation.

### In client projects.

ALTAI uses AfterImage to build training-data generation workflows for specific tasks. We define the task and example coverage, incorporate data review, and compare the trained model with a baseline on a separate test set.

### Run the first example.

The dry run calls no model. Real generation needs a provider key or a local model.

````
git clone https://github.com/altaidevorg/afterimage.git
cd afterimage
python -m pip install .

# Validate the configuration without calling a model
afterimage generate -c examples/configs/basic.yaml --dry-run

# Inspect supported export formats
afterimage export --list-formats
````

### Spec sheet

- **Runtime:** Python 3.11+
- **Interface:** YAML / Python / CLI
- **License:** Apache-2.0
- **Source:** [GitHub](https://github.com/altaidevorg/afterimage), [Documentation](https://afterimage.altai.dev/), [FunctionGemma demonstration](https://github.com/altaidevorg/functiongemma-afterimage-demo)
- **Scope:** A synthetic example is not proof of correctness. Plan a separate evaluation set. Provider terms apply separately.

### Write to us to use it in your organisation.

[Talk to us](/contact/?interest=AfterImage)

Source: https://altai.dev/open-source/afterimage/
