neulab/t-dig-train

T-DIG policy training data

The GRPO training data of the T-DIG generation policies, from T-DIG: Text-conditioned Diversity for Image Generation (under review). These are the exact datasets the paper's released generation policies were trained and validated on, one per T2I model: Qwen-Image-2512, Stable Diffusion 3.5 Large, Stable Diffusion 3.5 Medium and FLUX.2 [klein] 9B. Code: github.com/neulab/t-dig. Policies: neulab/t-dig. Test set: neulab/t-dig-test.

What a row is

A row gives the policy one base prompt and one axis, a request for what should differ across the images ("Monster creature type"), and asks it to write k = 4 prompts, one per image, that vary along the axis and otherwise keep the base prompt. Base prompts come from HPDv2; axes were proposed by a vision-language model that looked at four i.i.d. images of the prompt. An overall row has the axis overall and asks for variety across attributes the base prompt leaves open.

The four configurations hold the same base prompts, axes and order. They differ in the T2I model named in instruction_renderer, the word cap the instruction states, and the per-row thresholds, which are measured on that T2I model's own i.i.d. images.

Splits

Split	File	Rows	Contents
train	train.parquet	76,089	19,765 base prompts, 1,975 overall rows; four passes over the prompts in training order
validation	val.parquet	758	Every axis and overall row of 200 held-out base prompts, disjoint from training

Rows whose axis names something the images must exclude ("instead of", "rather than", an unpaired "without", a leading "Remove") are removed from both splits. The order of train is the training order: each pass holds one axis per prompt, rotating the axis slot from pass to pass.

Fields

• prompt: the chat message given to the policy, [{"role": "user", "content": <instruction>}]. instruction.txt shows the instruction of the first row.
• reward_model.ground_truth: a JSON string read by the reward:
  • base_prompt, axis: the base prompt and the axis, verbatim.
  • tpips_factor: the text that conditions the TPIPS diversity score; equal to axis.
  • k: images per set, 4.
  • align_thresh: mean VQAScore of the T2I model's four i.i.d. images of the base prompt against the base prompt. The fidelity penalty applies below it.
  • quality_thresh: mean HPSv3 of the same images. The quality penalty applies below it.
  • length_cap: word cap of each written prompt, the base prompt's word count plus 20 (21 to 70). The length penalty applies above it.
• extra_info: provenance and bookkeeping, including prompt_id (SHA-256 of the base prompt), row_id, index (training position), axis_kind (axis or overall), axis_slot and pass_number, cluster_id (HPDv2 prompt cluster), concise_factor (the elicitor's short name of the axis), source_renderer (the T2I model whose images the elicitor saw), instruction_version (tdig-v4-2), the thresholds again, align_thresh_images (images averaged) and length_cap.

data/<T2I model>/val-monitor.parquet is the trainer's validation file: per held-out prompt, its lowest-slot axis and its overall rows (224 rows, a subset of validation). data/<T2I model>/manifest.json lists row counts, settings, the build steps and the SHA-256 of every file; the trainer checks these hashes.

Prompt pool

prompts/curated-pool.json holds the 23,020 HPDv2 base prompts the paper started from: the prompts that passed both LLM filter passes, clustered by MiniLM embedding (sentence-transformers/all-MiniLM-L6-v2, cosine, average linkage) and cut at distance 0.35, keeping one prompt per cluster, drawn at random (seed 42) from the more central half of its members. Each entry in kept has the prompt's id (SHA-256 of the prompt), prompt, cluster_id and the filter pass's reason and variation_axes; cluster_summary lists every cluster's members. The training rows come from 20,000 of these prompts, 5,000 per T2I model. The code's iid.curated prepare --input prompts/curated-pool.json --allow-partial-input starts step (a) from this pool and reproduces the paper's i.i.d. images.

How it was built

The code repository's step (a) builds these files: HPDv2 prompt curation and clustering, four i.i.d. images per prompt per T2I model, axis elicitation, the axis export and a prompt-disjoint held-out split, then tdig.build_dataset --order rotated, tdig.iid_vqa attach, tdig.iid_quality attach and tdig.recap_dataset --relative-budget 20.

Training a policy with it

hf download neulab/t-dig-train --repo-type dataset --local-dir t-dig-train
T_DIG_SUBMIT="sbatch --nodes=2 --ntasks-per-node=1 --gpus-per-node=8 --time=72:00:00" \
$PY_TRAIN scripts/prepare_training_matrix.py --dataset t-dig-train/data/flux2-klein-9b \
  --out $T_DIG_RUNS/grpo-flux2-klein-9b --t2i-models flux2-klein-9b --sizes 9B \
  --tau 0.9 --lam 2.0 --learning-rate 2e-5 --metric vendi --length-weight 0.1 --quality-weight 0.05 --quality-margin 0 \
  --save-freq 50 --save-freq-early 25 --max-steps 2000 --validation offline \
  --track-metrics imagereward hpsv3 clip_score aesthetic --execute

These are the paper's settings. Then run scripts/preflight_training_matrix.py and scripts/submit_training_matrix.py as in the code repository's README.

License

CC BY 4.0. The base prompts are from HPDv2 (Apache-2.0) and remain under its license.

Citation

Citation information will be added after review.