loganbolton/vlm-verification-logs

VLM Verification Conversation Logs

Raw conversation logs from VLM solvers and judges across CharXiv and CountBench.

Every row is one model interaction: the image, the exact prompt(s) sent, the full model output, the extracted answer, and correctness. The verification runs include the verifier model's prompt, output and verdict.

Configs

config	rows	description
logs	644,093	one row per solver / verifier / agentic / rejection / self-consistency conversation
images	1,491	the unique source images (CharXiv + CountBench) as decodable Image features, keyed by (dataset, id)

Images are stored once in the images config and referenced from logs via image_path and (dataset, id) to avoid duplicating the same picture across the many logs that reference it.

from datasets import load_dataset

logs   = load_dataset("loganbolton/vlm-verification-logs", "logs",   split="train")
images = load_dataset("loganbolton/vlm-verification-logs", "images", split="train")

# join a log row to its image
img_by_key = {(r["dataset"], r["id"]): r["image"] for r in images}
row = logs[0]
pil = img_by_key[(row["dataset"], row["id"])]

Rows by experiment / record_type

experiment	rows		record_type	rows
verifier_grid	251,979		verify	297,056
rejection	259,547		rejection	259,547
agentic_vision	48,724		agentic	48,724
self_consistency	19,383		solver	19,383
countbench_intern	14,730		self_consistency	19,383
charxiv / charxiv_tier2	12,000 each
charxiv_tier3 / countbench*	2,946–6,000

record_type values

• solver — a model answering a benchmark question (model_output, extracted_answer, correct).
• verify — a solver answer judged by a verifier model (adds verifier_prompt, verifier_output, verifier_verdict). Includes the full N×N verifier grid.
• rejection — rejection-sampling runs; per-attempt records (extra.attempt, extra.accepted_attempt).
• agentic — zoom-tool agentic-vision runs (extra.turns, extra.crops, extra.n_crops); crops are box coordinates applied to the base image at inference time, so no crop bytes are stored separately.
• self_consistency — n-sample self-consistency; per-problem normalized answer keys in extra.keys / extra.key_correct (no full text output).

Columns (logs)

experiment, record_type, source_file, dataset, solver_model, verifier_model, id, image_path, image_exists, question, gold_answer, system_prompt (instruction template), task_prompt (rendered text prompt), rendered_prompt (full prompt incl. chat template / image tokens), model_output, extracted_answer, correct, verifier_prompt, verifier_rendered_prompt, verifier_output, verifier_verdict, generation_params (JSON), extra (family-specific fields, JSON), raw_metadata (JSON), raw_record (full original record, JSON).