Supervised-fine-tuning-only baseline (DS-MVTec 80.16 / VisA 64.78).
A Qwen2.5-VL-7B model fine-tuned for explainable industrial anomaly detection (IAD). Given a product image it produces a structured reasoning trace (<think>), a defect <location> and <type> (for anomalies), and a binary <answer>. Research artefact from the MSc thesis Reasoning-Enhanced Vision-Language Models for Explainable Industrial Anomaly Detection (TU Delft, 2026).
This is the plain starting point of the AnomalyThink family. The base model is fine-tuned directly on 6,000 teacher-written reasoning traces. There is no reinforcement-learning stage and no self-distillation loop. Use it as the reference point that the later models are measured against.
It is not the Keep-Correct-Revise (KCR) model. KCR is the recipe where an SFT plus GRPO policy generates its own traces, which are then kept when correct, corrected by a teacher when wrong, and revised when correct but weakly grounded, after which the base model is fine-tuned on that curated result. The KCR model is released separately as aacudad/AnomalyThink-Qwen2.5-VL-7B-KCR.
| Benchmark | Balanced accuracy |
|---|---|
| DS-MVTec (1,670) | 80.16% |
| VisA (2,141) | 64.78% |
Evaluated under a single common harness on the MMAD DS-MVTec and VisA subsets. This is the SFT-only headline before any RL stage.
Supervised fine-tuning on the AnomalyThink-6K corpus. 6,000 Gemini-2.5-Flash traces, vision encoder frozen, epoch 3 of 4.
The "AnomalyThink" reasoning traces were distilled from Gemini 2.5-Flash on Real-IAD images. Training data: aacudad/AnomalyThink (folder anomalythink_6k).
| Model | DS-MVTec | VisA | Note |
|---|---|---|---|
…-7B-KCR | 82.80 | 72.07 | thesis headline |
…-7B-SFT | 80.16 | 64.78 | this model |
…-7B-SFT-GRPO | 82.73 | 70.39 | this model plus GRPO |
…-7B-KCR-GRPO | 82.95 | 72.62 | research preview, not a thesis result |
import torch
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
model_id = "aacudad/AnomalyThink-Qwen2.5-VL-7B-SFT"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)
# Build the structured single-image IAD prompt + image, then generate.
Research on explainable IAD. Known limitations: the model can confidently hallucinate a defect on a normal part (false positive), and GRPO-lineage variants can over-predict the "Missing Parts" type. As the public DS-MVTec/VisA images may appear in VLM pretraining, absolute numbers should be read with that caveat.
@mastersthesis{acudad2026reasoning,
author = {Acudad, A.},
title = {Reasoning-Enhanced Vision-Language Models for Explainable Industrial Anomaly Detection},
school = {Delft University of Technology},
year = {2026},
type = {Master's thesis},
url = {https://resolver.tudelft.nl/uuid:65c62420-79c0-447f-b095-7fb11d4474fc}
}
Thesis: https://resolver.tudelft.nl/uuid:65c62420-79c0-447f-b095-7fb11d4474fc. Code and evaluation files: https://github.com/aacudad/IAD-VLMs.
Apache-2.0 (inherits the Qwen2.5-VL-7B base). Trained on Real-IAD (cite Real-IAD separately, images are not redistributed) with traces distilled from Gemini 2.5-Flash.
Supervised-fine-tuning-only baseline (DS-MVTec 80.16 / VisA 64.78).
A Qwen2.5-VL-7B model fine-tuned for explainable industrial anomaly detection (IAD). Given a product image it produces a structured reasoning trace (<think>), a defect <location> and <type> (for anomalies), and a binary <answer>. Research artefact from the MSc thesis Reasoning-Enhanced Vision-Language Models for Explainable Industrial Anomaly Detection (TU Delft, 2026).
This is the plain starting point of the AnomalyThink family. The base model is fine-tuned directly on 6,000 teacher-written reasoning traces. There is no reinforcement-learning stage and no self-distillation loop. Use it as the reference point that the later models are measured against.
It is not the Keep-Correct-Revise (KCR) model. KCR is the recipe where an SFT plus GRPO policy generates its own traces, which are then kept when correct, corrected by a teacher when wrong, and revised when correct but weakly grounded, after which the base model is fine-tuned on that curated result. The KCR model is released separately as aacudad/AnomalyThink-Qwen2.5-VL-7B-KCR.
| Benchmark | Balanced accuracy |
|---|---|
| DS-MVTec (1,670) | 80.16% |
| VisA (2,141) | 64.78% |
Evaluated under a single common harness on the MMAD DS-MVTec and VisA subsets. This is the SFT-only headline before any RL stage.
Supervised fine-tuning on the AnomalyThink-6K corpus. 6,000 Gemini-2.5-Flash traces, vision encoder frozen, epoch 3 of 4.
The "AnomalyThink" reasoning traces were distilled from Gemini 2.5-Flash on Real-IAD images. Training data: aacudad/AnomalyThink (folder anomalythink_6k).
| Model | DS-MVTec | VisA | Note |
|---|---|---|---|
…-7B-KCR | 82.80 | 72.07 | thesis headline |
…-7B-SFT | 80.16 | 64.78 | this model |
…-7B-SFT-GRPO | 82.73 | 70.39 | this model plus GRPO |
…-7B-KCR-GRPO | 82.95 | 72.62 | research preview, not a thesis result |
import torch
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
model_id = "aacudad/AnomalyThink-Qwen2.5-VL-7B-SFT"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(model_id)
# Build the structured single-image IAD prompt + image, then generate.
Research on explainable IAD. Known limitations: the model can confidently hallucinate a defect on a normal part (false positive), and GRPO-lineage variants can over-predict the "Missing Parts" type. As the public DS-MVTec/VisA images may appear in VLM pretraining, absolute numbers should be read with that caveat.
@mastersthesis{acudad2026reasoning,
author = {Acudad, A.},
title = {Reasoning-Enhanced Vision-Language Models for Explainable Industrial Anomaly Detection},
school = {Delft University of Technology},
year = {2026},
type = {Master's thesis},
url = {https://resolver.tudelft.nl/uuid:65c62420-79c0-447f-b095-7fb11d4474fc}
}
Thesis: https://resolver.tudelft.nl/uuid:65c62420-79c0-447f-b095-7fb11d4474fc. Code and evaluation files: https://github.com/aacudad/IAD-VLMs.
Apache-2.0 (inherits the Qwen2.5-VL-7B base). Trained on Real-IAD (cite Real-IAD separately, images are not redistributed) with traces distilled from Gemini 2.5-Flash.