ZoeyZou/generative-rlhf-training-data

Generative RLHF-V Online Interaction Training Data

This repository contains the exact 4,053-example multimodal preference mixture used by our online interaction training runs. The train split preserves the original training order through training_index.

Composition

Source	Examples
PKU-Alignment/align-anything	2,048
PKU-Alignment/BeaverTails-V	2,005
Total	4,053

The Align-Anything subset was sampled from text-image-to-text/new/train_40k.parquet using stratified sampling over the two response sources and the preferred label (12 buckets, seed 42). The BeaverTails-V subset contains every valid safe/unsafe pair produced by the training loader. The requested cap was 4,096 examples, but BeaverTails-V yielded 2,005 valid pairs, so the final interleaved mixture contains 4,053 examples. Every example includes one image.

Fields

• training_index: zero-based position in the actual training sequence.
• uid: stable identifier from the training loader.
• source: align-anything or beavertails-v.
• question: multimodal instruction or question.
• response_a, response_b: candidate responses.
• preferred: human-preferred label, A or B.
• image: the image used by the online training pipeline.

Provenance

{
  "cache_metadata": {
    "align_sample_mode": "stratified",
    "align_sample_seed": 42,
    "align_train_file": "text-image-to-text/new/train_40k.parquet",
    "allow_missing_images": false,
    "dataset_ids": {
      "align-anything": "PKU-Alignment/align-anything",
      "beavertails-v": "PKU-Alignment/BeaverTails-V"
    },
    "dataset_mix": [
      "align-anything",
      "beavertails-v"
    ],
    "max_align_examples": null,
    "max_beaver_examples": null,
    "max_examples": 4096,
    "version": 1
  },
  "cache_version": 1,
  "data_metrics": {
    "align_sample_bucket_count": 12.0,
    "align_sample_candidates": 39216.0,
    "align_sample_mode": "stratified",
    "align_sample_seed": 42.0,
    "align_sample_target": 2048.0,
    "dataset_mix": [
      "align-anything",
      "beavertails-v"
    ],
    "harmlessness_examples": 2005.0,
    "helpfulness_examples": 2048.0,
    "image_conditioned_examples": 4053,
    "image_materialized_count": 4053.0,
    "image_missing_count": 0.0,
    "images_per_example": 1.0,
    "prepared_examples_cache_hit": false,
    "prepared_examples_cache_key": "f22d8df1cdf3916a",
    "prepared_examples_cache_path": "/work/pi_d22939t_dartmouth_edu/f003kfd_dartmouth_edu/data_cache/interaction_prepared_examples/examples/f22d8df1cdf3916a.pkl",
    "prepared_examples_cache_version": 1,
    "source_counts": {
      "align-anything": 2048,
      "beavertails-v": 2005
    }
  },
  "exported_rows": 4053,
  "source_counts": {
    "align-anything": 2048,
    "beavertails-v": 2005
  }
}

Users must comply with the licenses and terms of the two upstream datasets.