GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
aacudad
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
aacudad
Model Tree
llava-hf/llava-onevision-qwen2-7b-si-hf
Input Modalities
Output Modalities
Supported Functionality
Metric is balanced accuracy, (TPR + TNR) / 2, in percent.
| Benchmark | Balanced accuracy |
|---|---|
| MMAD DS-MVTec (1,670 images) | 87.66 |
| MMAD VisA (2,141 images) | 72.38 |
Reference rows measured on the exact same harness:
| Model | DS-MVTec | VisA |
|---|---|---|
| LLaVA-OneVision-7B-SI base | 75.66 | 53.80 |
| IAD-R1 released checkpoint | 81.92 | 71.34 |
| AnomalyThink LLaVA SFT (the init for this run) | 85.91 | 68.26 |
| This model (SFT then GRPO) | 87.66 | 72.38 |
| AnomalyThink LLaVA KCR, first build (released weights) | 88.45 | 74.25 |
| AnomalyThink LLaVA KCR, corrected build (thesis row, epoch 2 / epoch 4) | 87.32 / 86.60 | 72.65 / 74.29 |
GRPO adds +1.75 on DS-MVTec and +4.12 on VisA over its own SFT initialisation. Most of that comes from recall. The SFT init misses a lot of defects and GRPO trades some precision for those misses.
Evaluation protocol. One shared harness for every row above. The DS-MVTec and VisA
subsets of MMAD, single image per prompt, the same instruction the model was trained on,
greedy decoding at temperature 0, at most 1024 new tokens, images capped at 262,144
pixels. Answers are parsed from the <answer> tag. Generation ran through vLLM 0.10.2,
which agreed with the plain transformers generate path on 99 percent of a probe set.
Nothing here is a re-scored or best-of-N number.
Strict scoring: a generation with no parsable verdict counts as wrong, as in the thesis. Six of the 2,141 VisA generations of the SFT then GRPO model have no verdict, which gives 72.38; over the 2,135 parsable outputs it is 72.58. No ordering changes.
The public LLaVA-OneVision training mixture (lmms-lab/LLaVA-OneVision-Data, config
vision_flan(filtered)) contains 426 rows whose id matches %MVTecAD%. The base model
has therefore seen MVTec-AD material during its own instruction tuning. Every DS-MVTec
number for any LLaVA-OneVision derived model carries that caveat, including the 87.66
above, and including the IAD-R1 row. We do not know how much of the gap is real
capability and how much is recall.
VisA is not affected. The same query over the mixture returns 0 rows for VisA. So the 72.58 on VisA is the clean number and it is the one to trust for a cross-model comparison. It is also where the win over IAD-R1 is smallest, at +1.24.
Two stages.
Stage 1, SFT. Base
llava-hf/llava-onevision-qwen2-7b-si-hf
fine-tuned on
anomalythink_6k/combined_6k_train.json,
6,000 AnomalyThink traces distilled from Gemini 2.5-Flash on Real-IAD images. Frozen
SigLIP vision tower, trainable projector and language model, learning rate 1e-5, cosine,
effective batch 32, bf16. The epoch-1 checkpoint of that run is published separately as
AnomalyThink-LLaVA-OneVision-7B-SFT
and is the exact initialisation for stage 2.
Stage 2, GRPO. Group Relative Policy Optimisation on
anomalythink_15k/c1_only_fixed/grpo_train.json,
4,236 Real-IAD prompts.
Epoch: this is epoch 1 of 2 (step 530). Epoch 2 scored 87.86 / 72.10, which is a wash, so epoch 1 is the released checkpoint. Note also that epoch 2 was restarted from this checkpoint after a server maintenance kill, so it has fresh optimiser state and a re-seeded dataloader. It is not a clean continuation.
One honest caveat on the objective. Our GRPO runs use beta = 0, so there is no KL term against the reference policy. IAD-R1's own code used beta = 0.04. Their paper shows a KL term in the objective but never states the value. This model therefore reproduces IAD-R1's recipe in structure, not in the exact objective, and the comparison should be read that way.
This model is not trained on a KCR corpus. The note still belongs here, because this checkpoint is the rollout source for that corpus.
KCR stands for keep, correct, rewrite. The corpus used by the sibling KCR model was built from LLaVA-OneVision's own rollouts, not from the Qwen rollouts used elsewhere in the thesis. This exact checkpoint generated 10,236 rollouts on Real-IAD training images at k = 8 and temperature 0.7. Each trace was bucketed into keep, needs correction, or needs rewrite, then scored by a Gemini-3-Flash faithfulness judge, then repaired where it failed, and finally sampled to a balanced 6,000.
All three arms of that corpus are published at
llava_kcr/:
sft_llava_A_kept.json, sft_llava_B_kept_corrected.json, and sft_llava_C_train.json.
Because the rollouts come from this backbone, the corpus is on-policy for LLaVA-OneVision, and that matters. The control, which is the Qwen derived KCR corpus trained on this same backbone, peaked at 87.70 on DS-MVTec and 73.52 on VisA in different epochs. The native corpus reaches 88.45 and 74.25 in one checkpoint.
The weights and configs in this repo were written by transformers 5.0.0. Loading, the processor, and the full evaluation were verified under transformers 4.57.1, both through the plain HF generate path and through vLLM 0.10.2.
This checkpoint is the one the rope trap actually bit, so read this if you fork it.
Transformers 5.0 writes the rope settings under text_config.rope_parameters.
Transformers 4.x does not read that key and silently falls back to rope_theta = 10000,
which is 100 times too small. The model stays perfectly fluent but goes blind. It
answered "no" to almost everything and scored 27.5 on DS-MVTec, which looks like a
collapsed training run rather than a loading bug. It cost us a day.
The config in this repo carries both forms, so it loads correctly on 4.x and on 5.x:
json
"text_config": {"rope_parameters": { "rope_theta": 1000000.0, "rope_type": "default" },"rope_theta": 1000000.0}
If you re-save this model from transformers 5.0, check that the plain rope_theta key
survives before you evaluate on 4.x.
python
import torchfrom PIL import Imagefrom transformers import AutoProcessor, LlavaOnevisionForConditionalGenerationrepo = "aacudad/AnomalyThink-LLaVA-OneVision-7B-SFT-GRPO"model = LlavaOnevisionForConditionalGeneration.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")processor = AutoProcessor.from_pretrained(repo)image = Image.open("part.png").convert("RGB")product = "tile" # the product category, it goes into the promptquestion = (f"Analyze the provided image of the {product}. ""Determine if there are any anomalies present. ""If an anomaly is detected, specify its type and location, ""and provide a detailed reasoning for your conclusion.")messages = [{"role": "user", "content": [{"type": "image"},{"type": "text", "text": question},]}]prompt = processor.apply_chat_template(messages, add_generation_prompt=True)inputs = processor(images=image, text=prompt, return_tensors="pt").to(model.device, torch.bfloat16)out = model.generate(**inputs, max_new_tokens=1024, do_sample=False)print(processor.decode(out[0][inputs["input_ids"].shape[1]:],skip_special_tokens=True))
Use this exact instruction. The model was trained on it and it degrades on a different phrasing.
Expected output on a defective part:
markdown
<think>I am inspecting a tile with a speckled, grayish-white surface. ... In the center of thetile, I detect a triangular, translucent plastic fragment. ...</think><location>center</location><type>Contamination</type><answer>Yes</answer>
On a normal part the model emits <think> and then <answer>No</answer>, with no
<location> or <type> tag.
Research on explainable industrial anomaly detection. This is a thesis artefact, not a production inspection system.
Known limitations:
bibtex
@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. The training data is at aacudad/AnomalyThink.
Apache-2.0, inherited from the LLaVA-OneVision-7B-SI base. Trained on Real-IAD images, which are not redistributed here, so cite Real-IAD separately. Reasoning traces were distilled from Gemini 2.5-Flash.