Skip to main content

First Train Pairs

train-pairs is AutoPipeline's post-processing entry point. It converts intermediate artifacts into pairwise data that can be consumed by a training pipeline.

Accepted inputs​

Two input families are supported:

  • grouped results produced by annotation
  • pairwise grouped results produced by eval or another judge-based workflow

The CLI inspects the results field and automatically detects:

  • group mode
  • judge mode

Minimal command​

cd <PROJECT_ROOT>
autopipeline train-pairs \
  --tasks subject_add \
  --input-dir <PROJECT_ROOT>/data/c_annotated_group_data \
  --output-dir <PROJECT_ROOT>/data/d_train_data \
  --mode auto

Common parameters​

  • --tasks Comma-separated task list. Multiple tasks can be processed in one run.
  • --prompts-num Maximum number of prompts or groups to process per task.
  • --mode auto, group, or judge.
  • --filt-out-strategy head_tail or three_tiers. Only applies to group mode.
  • --thresholds-config-file Threshold config used for tiering and filtering in group mode.

Output files​

Each task produces one JSON file:

<output-dir>/<task>.json

Simplified output example​

[
  {
    "edited_image_paths": [
      "images/edited/model_a/0001.png",
      "images/edited/model_b/0001.png"
    ],
    "instruction": "Add a red balloon to the left side of the boy.",
    "source_image_path": "images/source/0001.png",
    "gpt_response": "```json\n{\n    \"winner\": \"Image A\"\n}\n```"
  }
]

group mode vs judge mode​

group mode​

This mode consumes grouped annotation results from human-centric or object-centric pipelines.

The conversion logic will:

  • clean the candidate set
  • compute z-scores
  • assign tiers according to area-specific configs
  • generate chosen/rejected pairs

judge mode​

This mode consumes pairwise judge outputs.

It reads the winner directly and converts each judged pair into a training sample.

When to use it​

If your goal is not just to inspect evaluation artifacts but to continue into training, this is the correct next step.