pdsdpo/synthalign-length-adjusted-9k

SynthAlign Length-Adjusted 9K

This repository releases 9,035 unique image–question preference pairs from the full-9K SynthAlign paper. The images and four-response candidate pool from ImageReward, but the chosen/rejected endpoints are selected using length-adjusted ArmoRM scores rather than raw scores.

Length-adjusted does not mean length-matched. Responses are not rewritten, shortened, or padded. The adjustment changes which existing candidate answers are selected. These are reward-guided proxy labels, not human-verified labels or direct verification of visual correctness.

Project: SynthAlign · Source assets: SynthAlign v1.0 · Project paper

Contents

Item	Value
Unique preference pairs	9,035
Distinct question texts	9,017
Candidate responses per image–question	4
Images	Original released ImageReward-selected synthetic JPEGs
Response ranker	RLHFlow/ArmoRM-Llama3-8B-v0.1
Ranker input	Question and answer only; no image pixels
Calibration	5 question-text-grouped cross-fitting folds
Split	train only; no held-out validation/test split
Batch-padding duplicates in release split	0

The upstream release contains 9,036 records.

Response pool and scoring

Four image-conditioned candidate answers were generated using:

1. llava-hf/llava-v1.6-mistral-7b-hf
2. llava-hf/llava-v1.6-vicuna-13b-hf
3. OpenGVLab/InternVL2_5-8B
4. OpenGVLab/Mini-InternVL-Chat-4B-V1-5

Pinned model revisions, deterministic decoding settings, retry policies, the source dataset revision, and original artifact hashes are recorded in construction.json. ArmoRM scored the four candidates in BF16 using the associated question as the user message and the candidate answer as the assistant message. Image markers were removed from the question for generation/ranking; image pixels were supplied to the response generators but not to ArmoRM. Response length is measured with the pinned ArmoRM tokenizer, with add_special_tokens=False.

Length adjustment

For candidate answer (y_j) to question (q_i), the adjusted score is

[ \widetilde r_{ij}=r_{ij}-\beta_{-k(i)}\ln L_{ij}, ]

where (r_{ij}) is the raw ArmoRM score and (L_{ij}) is the answer's token count. The coefficient for each held-out fold is estimated on the other four folds by within-question-centered ordinary least squares:

[ \beta_{-k}=\frac{\sum_{i\notin k,j}(\ln L_{ij}-\overline{\ln L}i) (r{ij}-\bar r_i)}{\sum_{i\notin k,j}(\ln L_{ij}-\overline{\ln L}_i)^2}. ]

Duplicate question texts share a fold. Fold assignment uses seed 42. Within each image–question group, the maximum adjusted score is chosen and the minimum is rejected; exact endpoint ties are broken by the fixed teacher order above. All fold assignments and coefficients are in crossfit_calibration.json.

This is dataset-internal cross-fitting, not independent external validation. No benchmark labels are used. The adjustment removes a fitted linear association with log answer length; it does not establish or eliminate a causal verbosity bias, nor does it guarantee equal-length or visually correct answers.

Dataset fields and use

from datasets import load_dataset

ds = load_dataset("pdsdpo/synthalign-length-adjusted-9k", split="train")
assert len(ds) == 9035
example = ds[0]
image = example["image"]       # decoded PIL image
question = example["prompt"]
chosen = example["chosen"]
rejected = example["rejected"]

The 19 Parquet shards embed the original image bytes and require no access to the authors' workstation. Core columns are id, image, prompt, chosen, and rejected. source_question preserves the original training question, including its image-marker convention. Additional columns contain category, selected teacher IDs, token counts, raw and adjusted endpoint scores, adjusted margin, cross-fit fold/coefficient, image SHA-256, and whether the ordered pair differs from raw-score max/min selection.

For conversation-format DPO data with local image files:

pip install datasets pillow
python export_dpo.py --output ./synthalign-length-adjusted-9k

This writes images/, dpo_unique.json, and dpo_pairs.jsonl. The latter is also provided directly in this repository; its relative image paths refer to the images materialized by the export script. Do not treat those paths as separate image files already stored in the repository: the canonical images are embedded in Parquet.

To reproduce the original experiment's 9,040 training exposures, add --include-padding. This appends the five explicitly listed repeats in training_padding.json, while keeping dpo_unique.json at 9,035 rows. These repeats are an implementation detail of the original effective batch size of eight, not new data.

Descriptive checks

Measurement	Value
Mean chosen response tokens	252.57
Mean rejected response tokens	238.95
Ordered pairs changed from raw-score max/min	5,499 / 9,035
Positive raw reward margin after adjustment	7,759
Negative raw reward margin after adjustment	1,176
Zero raw reward margin after adjustment	100

Every selected pair has a positive adjusted margin. A negative raw margin is possible and intentional because selection uses adjusted, not raw, scores. These counts describe dataset construction; they are not downstream evaluation results or evidence that one condition outperforms another.

pair_provenance.json contains endpoint hashes and selection metadata. validation.json records the release checks, per-shard counts/hashes, and the earlier training-tokenization audit. The complete four-answer candidate pool is not included in this release.

Limitations and intended use

This research dataset supports controlled preference-tuning experiments on synthetic images. Text-only ArmoRM may prefer plausible but visually incorrect answers, and residual length/style or teacher-specific effects can remain. Synthetic images and generated answers may contain artifacts, inaccuracies, stereotypes, or hallucinations. Neither the labels nor the adjustment should be described as human alignment or a visual-grounding guarantee.

Use a separate evaluation set. This release contains no new human annotations, and has not been certified for safety-critical deployment. Image-generation prompts and response questions are different concepts; the retained questions are those used in the original experiment.

Source licensing and attribution

The original SynthAlign v1.0 dataset card declares Apache-2.0. This derived release reuses its synthetic images/questions and adds newly generated responses and preference-selection metadata. Users must also review the upstream response-generator and reward-model licenses and terms linked by their model IDs. This card does not assert that a single blanket license overrides those upstream terms or grant additional rights to third-party components.