AnomalyThink-MMAD: reasoning traces for the MMAD benchmark

aacudad/AnomalyThink-MMAD holds one structured reasoning trace for 8,293 of the 8,366 images of the MMAD benchmark (DS-MVTec, VisA, GoodsAD, MVTec-LOCO), written with the same teacher recipe as the AnomalyThink corpus of the MSc thesis Reasoning-Enhanced Vision-Language Models for Explainable Industrial Anomaly Detection (TU Delft, 2026). Each trace is a single-image inspection: a <think> block, and for anomalies a <location> on a 3x3 grid and a <type>, then a yes/no <answer>. The MMAD annotations (mask, defect type, location, appearance, effect) were given to the teacher, so the traces are grounded in the benchmark's own ground truth.

Read this before training on it. MMAD is an evaluation benchmark. A model fine-tuned on any part of this corpus and then scored on MMAD has seen the benchmark's products, cameras and defect vocabulary, and its MMAD numbers are not comparable with models that have not. We used the corpus to measure exactly that effect (section Leakage test below). No thesis model was trained on any MMAD image, and the checkpoints fine-tuned on this corpus are deliberately not released.

Important: images are NOT included

Every example references an MMAD image by relative path (MMAD/DS-MVTec/bottle/image/broken_large/000.png), which is the key of the image in MMAD's mmad.json. Obtain MMAD from its official source (https://huggingface.co/datasets/jiang-cc/MMAD, CC BY-NC-SA 4.0) and place it so the paths resolve. This dataset inherits MMAD's non-commercial licence.

How the traces were written

• Teacher: gemini-3.6-flash (Vertex AI), thinking level MINIMAL, JSON output, batches of 10.
• System prompt: inspector_prompt_test_v2 of the thesis (six-phase inspection, XML tags, hard rules, auto-reject rules), user prompt of the v4 Real-IAD generator.
• Inputs per anomalous image: the image, the same image with the MMAD ground-truth mask as a red overlay, and a normal reference image of the same product taken from MMAD's similar_templates. Normal images: the image only.
• Hints: the MMAD multiple-choice answers for defect type, location, appearance and effect were given to the teacher as internal labels and are stored in hints of traces_8293.jsonl.
• Output: <think>...</think> then, for anomalies, <location> on a 3x3 grid, <type> from the thesis defect vocabulary, and <answer>Yes</answer>, otherwise <answer>No</answer>. Mean think length 113 words (anomalous) and 127 words (normal).
• Excluded: 69 images without a verdict question in MMAD and 4 MVTec-LOCO anomalies without a mask.

Files

file	rows	content
mmad_all_8293.json	8,293	ShareGPT (LLaMA-Factory) records, all traces
traces_8293.jsonl	8,293	one record per image with hints, dataset, product, is_anomaly, teacher, and the split membership of both runs
splits/run1_unbalanced_train_1600.json / _test_6693.json	1,600 / 6,693	19.3 % of every (dataset, product, label) stratum, MMAD's own label ratio
splits/run2_balanced_train_1600.json / _test_6693.json	1,600 / 6,693	800 anomalous + 800 normal, at most 60 % of a product's images of either label, 6,521 test images shared with run 1
splits/one_per_product_and_defect_144.json	144	one image per (dataset, product, defect folder), the OmniAD "one example per category" reading
splits/one_per_product_39.json	39	one image per product, the other reading
splits/split_keys.json		the key lists of every split
RESULTS.md		the experiment: recipe, epoch tables, sensitivity/specificity, memorisation checks
system_prompt_inspector_v2.txt		the exact system prompt the teacher was given (the user prompt is built in the generator)

Leakage test: what training on part of MMAD buys

Several published VLM anomaly detectors put MMAD images into their training data (OmniAD v1: "one example per category from MMAD" for both SFT and GRPO; AnomalyR1: 600 images from the four MMAD source datasets) and report MMAD scores. To measure the effect we fine-tuned Qwen2.5-VL-7B with the thesis SFT recipe on 1,600 of these traces (two splits: MMAD's own label ratio, and 800/800) and scored the other 6,693 images of the same products with the thesis harness (strict balanced accuracy, unparsed = wrong, zero-shot, single image). Full tables in RESULTS.md.

Strict balanced accuracy on unseen images of the same products:

Subset	Qwen2.5-VL-7B base	1,600 MMAD traces, balanced, best epoch	thesis SFT on 6,000 Real-IAD traces (no MMAD)
DS-MVTec	69.7	79.2	80.2
VisA	53.8	67.2	64.8
GoodsAD	51.0	61.1	n/a
MVTec-LOCO	50.5	54.7	n/a

Scoring the same checkpoint on its own training images gives only 1 to 2 points more than on the held-out images, so the gain is familiarity with the benchmark's products, not memorisation. The thesis KCR model, which never saw an MMAD image, scores 82.4 / 73.4 / 58.4 / 50.2 on the same held-out keys (DS-MVTec / VisA / GoodsAD / MVTec-LOCO); pooled over the four subsets that is 66.0 against 65.4 for the best benchmark-trained run and 61.4 zero-shot for OmniAD-7B.

Reproduce the traces and the experiment

Code and the exact prompt are in the GitHub repository, folder scripts/06_mmad:

• generator: https://github.com/aacudad/IAD-VLMs/blob/main/scripts/06_mmad/generate_mmad_traces_v4.py (the user prompt, the three-image input and the hint lines are built inside format_anomaly_prompt)
• system prompt, the exact file the run used, also shipped here as system_prompt_inspector_v2.txt: https://github.com/aacudad/IAD-VLMs/blob/main/scripts/06_mmad/inspector_prompt_test_v2_mmad_run.txt
• split, SFT recipe, evaluation and scoring: compile_split.py, compile_split_bal.py, sft_mmad_train1600*.yaml, eval_heldout_vllm.py, score_heldout.py in the same folder, write-up in RESULTS.md
• per-checkpoint evaluation files: https://github.com/aacudad/IAD-VLMs/tree/main/results/mmad

export WORK_DIR=/path/to/your/clone   # MMAD under $WORK_DIR/reasoning_traces_gen/data/MMAD
python generate_mmad_traces_v4.py --shard_id 0 --total_shards 8   # one process per shard, Vertex AI credentials in GOOGLE_APPLICATION_CREDENTIALS
python compile_split.py && python compile_split_bal.py

Citation

Cite the thesis and MMAD. Code: https://github.com/aacudad/IAD-VLMs (scripts/06_mmad).