Super-squash branch 'main' using huggingface_hub

Co-authored-by: ruke1ire <ruke1ire@users.noreply.huggingface.co>

.gitattributes

@@ -0,0 +1,37 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+sample_image.png filter=lfs diff=lfs merge=lfs -text
+lfm2_vl_450m_metrics.png filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,71 @@
+LFM Open License v1.0 
+
+TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+1. Definitions.
+
+"License" shall mean the terms and conditions for use, reproduction, and distribution as defined by this document.
+
+"Licensor" shall mean Liquid AI, Inc.
+
+"Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.
+
+"You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License.
+
+"Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
+
+"Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
+
+"Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work.
+
+"Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.
+
+"Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
+
+"Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.
+
+"Commercial Use" shall mean any use of the Work for direct or indirect commercial advantage or monetary compensation.
+
+"Qualified Non-Profit Organization" shall mean a Legal Entity that is organized and operated exclusively for religious, charitable, scientific, testing for public safety, literary, or educational purposes, and which is exempt from federal income tax under Section 501(c)(3) of the United States Internal Revenue Code of 1986, as amended, or any equivalent non-profit or charitable organization in a foreign jurisdiction.
+
+"Non-Commercial or Research Purposes" shall mean purposes that do not involve any use of the Work or a Derivative Work for Commercial Use.
+
+"Threshold" shall mean annual revenue of 10 million United States dollars ($10,000,000) or more.
+
+2. Grant of Copyright License. Subject to the terms and conditions of this License, including the Commercial Use limitation set forth in Section 5, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
+
+3. Grant of Patent License. Subject to the terms and conditions of this License, including the Commercial Use limitation set forth in Section 5, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.
+
+4. Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
+
+(a) You must give any other recipients of the Work or Derivative Works a copy of this License; and
+
+(b) You must cause any modified files to carry prominent notices stating that You changed the files; and
+
+(c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and
+
+(d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License.
+
+You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.
+
+5. Commercial Use Limitation.
+
+(a) The rights granted under this License for Commercial Use are conditioned upon You or Your Legal Entity not exceeding the Threshold.
+
+(b) Any Commercial Use of the Work or a Derivative Work by a Legal Entity that exceeds the Threshold is not licensed under this Agreement.
+
+(c) The Threshold shall not apply to a Qualified Non-Profit Organization's use of the Work or a Derivative Work for Non-Commercial or Research Purposes.
+
+6. Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.
+
+7. Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except for the reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.
+
+8. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.
+
+9. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
+
+10. Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
+
+11. Termination. This License will terminate automatically and immediately if You fail to comply with any of its terms and conditions. Upon termination, You must cease all use of the Work and any Derivative Works and delete all copies in Your possession.
+
+END OF TERMS AND CONDITIONS

README.md

@@ -0,0 +1,216 @@
+---
+library_name: transformers
+license: other
+license_name: lfm1.0
+license_link: LICENSE
+language:
+- en
+pipeline_tag: image-text-to-text
+tags:
+- liquid
+- lfm2.5
+- lfm2
+- edge
+- vision
+base_model: LiquidAI/LFM2.5-VL-450M
+---
+<center>
+<div style="text-align: center;">
+  <img 
+    src="https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/2b08LKpev0DNEk6DlnWkY.png" 
+    alt="Liquid AI"
+    style="width: 100%; max-width: 100%; height: auto; display: inline-block; margin-bottom: 0.5em; margin-top: 0.5em;"
+  />
+</div>
+<div style="display: flex; justify-content: center; gap: 0.5em;">
+<a href="https://playground.liquid.ai/chat?model=lfm2.5-vl-450m"><strong>Try LFM</strong></a> • <a href="https://docs.liquid.ai/lfm/getting-started/welcome"><strong>Docs</strong></a> • <a href="https://leap.liquid.ai/"><strong>LEAP</strong></a> • <a href="https://discord.com/invite/liquid-ai"><strong>Discord</strong></a>
+</div>
+</center>
+
+<br>
+
+# LFM2.5-VL-450M-Extract
+
+**LFM2.5-VL-450M-Extract** extracts user-defined fields from images and returns them as **JSON**. It is Liquid AI's first vision model in the [Liquid Nanos](https://huggingface.co/collections/LiquidAI/liquid-nanos) collection—compact, task-specific models built for production workflows—and extends the Extract family alongside [LFM2-350M-Extract](https://huggingface.co/LiquidAI/LFM2-350M-Extract) for text documents.
+
+You specify what to extract as a YAML field list in the system prompt, and the model returns a JSON object with those fields. Structured outputs integrate cleanly with rule-based systems and downstream pipelines. Use it out of the box or fine-tune for domain-specific extraction.
+
+## ⚙️ How it works
+
+You specify what to extract as a YAML field list in the system prompt, and the model returns a JSON object with those fields. Structured outputs integrate cleanly with rule-based systems and downstream pipelines. Use it out of the box or fine-tune for domain-specific extraction.
+
+- **System prompt**:
+
+```yaml
+wood_color: The overall coloration of the wood surface
+wood_texture: The tactile quality of the wood surface 
+wood_pattern: The partern types visible on the wood surface
+```
+
+- **User prompt**:
+<img src="https://huggingface.co/LiquidAI/LFM2.5-VL-450M-Extract/resolve/main/sample_image.png" width="300">
+
+- **Output**:
+```yaml
+{
+  "wood_color": "light to medium brown",
+  "wood_texture": "smooth with visible grain",
+  "wood_pattern": "parallel, irregular, wavy"
+}
+```
+
+Our model supports the enum feature, which lets you provide a list of possible choices alongside the field description as follows, and the model will return one of the listed values as its answer. 
+
+- **System prompt**:
+
+```yaml
+wood_color: The overall coloration of the wood surface, such as blue, red, or light tan
+wood_texture: The tactile quality of the wood surface, select from smooth, rough, or grainy
+wood_pattern: The partern types visible on the wood surface, e.g., straight, wavy, or curly
+```
+
+## 🌟 Use cases
+
+- Detecting safety-critical events in images (e.g. fallen person, fire, leakage) to trigger automated safety systems.
+- Collecting statistical information about objects across video frames for analytics pipelines.
+- Auto-tag product images with structured attributes for Retail/E-commerce.
+
+## 📄 Model details
+
+| Property | Detail |
+|---|---:|
+| **Parameters (LM only)** | 350M |
+| **Vision encoder** | SigLIP2 (~100M, [SigLIP-2 paper](https://arxiv.org/abs/2502.14786)) |
+| **Backbone layers** | hybrid conv+attention |
+| **Image input** | Single image, dynamic resolution |
+| **Context** | 128,000 tokens |
+| **Vocab size** | 65,536 (text) |
+| **Precision** | bfloat16 |
+| **License** | LFM Open License v1.0 |
+
+## 📊 Performance
+
+We evaluated LFM2.5-VL-450M-Extract on a 2,000-sample benchmark of
+`(image, schema, JSON)` triples, with reference labels generated by an
+ensemble of frontier multimodal models. Predictions are scored on the
+following three dimensions:
+
+- **JSON Validity** — share of samples producing strict-parseable JSON
+- **Schema Consistency F1 Score** — set-level F1 over predicted vs requested field names, macro-averaged across samples
+- **VLM Judge Score** — match against the image directly, judged by a separate vision model ([Qwen/Qwen3.5-35B-A3B](https://huggingface.co/Qwen/Qwen3.5-35B-A3B))
+
+<img src="https://huggingface.co/LiquidAI/LFM2.5-VL-450M-Extract/resolve/main/lfm2_vl_450m_metrics.png" width="800">
+
+| Model | Params | JSON Validity | F1 Score | VLM Judge Score |
+|---|---:|---:|---:|---:|
+| **LFM2.5-VL-450M-Extract** | **0.45B** | **98.9** | **98.8** | **84.5** |
+| LFM2.5-VL-450M | 0.45B | 97.7 | 93.5 | 73.4 |
+| SmolVLM-500M-Instruct | 0.51B | 33.0 | 26.6 | 12.2 |
+| FastVLM-0.5B | 0.76B | 22.5 | 19.3 | 16.3 |
+| Qwen3.5-0.8B | 0.87B | 96.4 | 96.3 | 82.3 |
+| InternVL3_5-1B | 1.06B | 98.0 | 96.5 | 80.7 |
+| MiniCPM-V-4.6 | 1.30B | 61.8 | 60.4 | 57.5 |
+| *(ref) InternVL3_5-2B* | 2.35B | 99.6 | 99.2 | 87.7 |
+| *(ref) Qwen3.5-2B* | 2.27B | 97.9 | 97.7 | 89.7 |
+| *(ref) gemma-4-E2B-it* | 2.3B | 97.4 | 97.1 | 84.4 |
+
+LFM2-VL-450M-Extract outperforms similarly-sized (sub-1B) open-source VLMs on this benchmark and is competitive with models 4× its size.
+
+**Reproducing these numbers**: The full evaluation pipeline, which includes extraction, VLM judging, and metric aggregation, is bundled in this repository under `model_eval/`. Setup, configuration, and run instructions are in the folder's [`README`](./model_eval/README.md).
+
+**Scope**: These numbers characterize the model on the input/output form it is designed for: a single input image, a YAML field list as the schema, and a flat JSON object as the output. Performance is not expected to transfer to largely different tasks, e.g. multi-image reasoning or free-form VQA.
+
+<!-- > Generic instruction-tuned VLMs (SmolVLM, moondream) cannot perform
+> schema-based extraction zero-shot regardless of prompt strategy.
+> Under the most permissive prompt setups, they either:
+>   - produce free-form captions ignoring the JSON instruction, or
+>   - produce valid-shaped JSON but echo the schema descriptions or
+>     few-shot example values as field values (zero faithfulness to
+>     the image).
+>
+> LFM2-VL-Extract's task-specific training is what enables strict-JSON
+> output with faithful, image-grounded values in a single zero-shot
+> call — no few-shot examples, no grammar constraints, no inference
+> wrappers. -->
+
+
+The full evaluation pipeline, which includes extraction, LLM/VLM judging, and
+metric aggregation, is included in this repository under `model_eval/`. Usage details are in the folder's README.
+
+## 🏃 How to run
+
+You can run LFM2.5-VL-450M-Extract with Hugging Face [`transformers`](https://github.com/huggingface/transformers) v5.1 or newer:
+
+```bash
+pip install transformers pillow
+```
+
+```python
+from transformers import AutoProcessor, AutoModelForImageTextToText
+from transformers.image_utils import load_image
+
+model_id = "LiquidAI/LFM2.5-VL-450M-Extract"
+model = AutoModelForImageTextToText.from_pretrained(
+    model_id,
+    device_map="auto",
+    dtype="bfloat16",
+    trust_remote_code=True,
+)
+processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
+
+image = load_image("https://huggingface.co/LiquidAI/LFM2.5-VL-450M-Extract/resolve/main/sample_image.png")
+
+fields_yaml = """wood_color: The overall coloration of the wood surface
+wood_texture: The tactile quality of the wood surface
+wood_pattern: The pattern types visible on the wood surface"""
+
+system_prompt = f"""Extract the following from the image:
+
+{fields_yaml}
+
+Respond with only a JSON object. Do not include any text outside the JSON."""
+
+conversation = [
+    {"role": "system", "content": system_prompt},
+    {"role": "user",   "content": [{"type": "image", "image": image}]},
+]
+
+inputs = processor.apply_chat_template(
+    conversation,
+    add_generation_prompt=True,
+    return_tensors="pt",
+    return_dict=True,
+    tokenize=True,
+).to(model.device)
+
+outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
+response = processor.batch_decode(
+    outputs[:, inputs["input_ids"].shape[1]:],
+    skip_special_tokens=True,
+)[0]
+print(response)
+# {
+#   "wood_color": "light to medium brown",
+#   "wood_texture": "smooth with visible grain",
+#   "wood_pattern": "parallel, irregular, wavy"
+# }
+```
+
+> [!WARNING]
+> The model is intended for single-turn conversations. We recommend using greedy decoding (`temperature=0`).
+
+## 📬 Contact
+
+- Got questions or want to connect? [Join our Discord community](https://discord.com/invite/liquid-ai)
+- If you are interested in custom solutions with edge deployment, please contact [our sales team](https://www.liquid.ai/contact).
+
+## Citation
+
+```bibtex
+@article{liquidai2025lfm2,
+ title={LFM2 Technical Report},
+ author={Liquid AI},
+ journal={arXiv preprint arXiv:2511.23404},
+ year={2025}
+}
+```

chat_template.jinja

@@ -0,0 +1,92 @@
+{{- bos_token -}}
+{%- set keep_past_thinking = keep_past_thinking | default(false) -%}
+
+{%- macro format_arg_value(arg_value) -%}
+    {%- if arg_value is string -%}
+        {{- '"' + arg_value + '"' -}}
+    {%- elif arg_value is mapping -%}
+        {{- arg_value | tojson -}}
+    {%- else -%}
+        {{- arg_value | string -}}
+    {%- endif -%}
+{%- endmacro -%}
+
+{%- macro parse_content(content) -%}
+    {%- if content is string -%}
+        {{- content -}}
+    {%- else -%}
+        {%- set _ns = namespace(result="") -%}
+        {%- for item in content -%}
+            {%- if item.type == "image" -%}
+                {%- set _ns.result = _ns.result + "<image>" -%}
+            {%- elif item.type == "text" -%}
+                {%- set _ns.result = _ns.result + item.text -%}
+            {%- else -%}
+                {%- set _ns.result = _ns.result + item | tojson -%}
+            {%- endif -%}
+        {%- endfor -%}
+        {{- _ns.result -}}
+    {%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_tool_calls(tool_calls) -%}
+    {%- set tool_calls_ns = namespace(tool_calls=[]) -%}
+    {%- for tool_call in tool_calls -%}
+        {%- set func_name = tool_call.function.name -%}
+        {%- set func_args = tool_call.function.arguments -%}
+        {%- set args_ns = namespace(arg_strings=[]) -%}
+        {%- for arg_name, arg_value in func_args.items() -%}
+            {%- set args_ns.arg_strings = args_ns.arg_strings + [arg_name + "=" + format_arg_value(arg_value)] -%}
+        {%- endfor -%}
+        {%- set tool_calls_ns.tool_calls = tool_calls_ns.tool_calls + [func_name + "(" + (args_ns.arg_strings | join(", ")) + ")"] -%}
+    {%- endfor -%}
+    {{- "<|tool_call_start|>[" + (tool_calls_ns.tool_calls | join(", ")) + "]<|tool_call_end|>" -}}
+{%- endmacro -%}
+
+{%- set ns = namespace(system_prompt="", last_assistant_index=-1) -%}
+{%- if messages[0].role == "system" -%}
+    {%- if messages[0].content is defined -%}
+        {%- set ns.system_prompt = parse_content(messages[0].content) -%}
+    {%- endif -%}
+    {%- set messages = messages[1:] -%}
+{%- endif -%}
+{%- if tools -%}
+    {%- set ns.system_prompt = ns.system_prompt + ("\n\n" if ns.system_prompt else "") + "Today's date: " + strftime_now("%Y-%m-%d") + "\n\nList of tools: " + (tools | tojson) -%}
+{%- endif -%}
+{%- if ns.system_prompt -%}
+    {{- "<|im_start|>system\n" + ns.system_prompt + "<|im_end|>\n" -}}
+{%- endif -%}
+{%- for message in messages -%}
+    {%- if message.role == "assistant" -%}
+        {%- set ns.last_assistant_index = loop.index0 -%}
+    {%- endif -%}
+{%- endfor -%}
+{%- for message in messages -%}
+    {{- "<|im_start|>" + message.role + "\n" -}}
+    {%- if message.role == "assistant" -%}
+        {%- generation -%}
+        {%- if message.thinking is defined and (keep_past_thinking or loop.index0 == ns.last_assistant_index) -%}
+            {{- "<think>" + message.thinking + "</think>" -}}
+        {%- endif -%}
+        {%- if message.tool_calls is defined -%}
+            {{- render_tool_calls(message.tool_calls) -}}
+        {%- endif -%}
+        {%- if message.content is defined -%}
+            {%- set content = parse_content(message.content) -%}
+            {%- if not keep_past_thinking and loop.index0 != ns.last_assistant_index -%}
+                {%- if "</think>" in content -%}
+                    {%- set content = content.split("</think>")[-1] | trim -%}
+                {%- endif -%}
+            {%- endif -%}
+            {{- content + ("" if (continue_final_message and loop.last) else "<|im_end|>\n") -}}
+        {%- endif -%}
+        {%- endgeneration -%}
+    {%- else %}
+        {%- if message.content is defined -%}
+            {{- parse_content(message.content) + "<|im_end|>\n" -}}
+        {%- endif -%}
+    {%- endif %}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    {{- "<|im_start|>assistant\n" -}}
+{%- endif -%}

config.json

@@ -0,0 +1,104 @@
+{
+  "architectures": [
+    "Lfm2VlForConditionalGeneration"
+  ],
+  "bos_token_id": 1,
+  "do_image_splitting": true,
+  "downsample_factor": 2,
+  "dtype": "bfloat16",
+  "encoder_patch_size": 16,
+  "eos_token_id": 7,
+  "image_token_id": 396,
+  "max_image_tokens": 256,
+  "max_pixels_tolerance": 2.0,
+  "max_tiles": 10,
+  "min_image_tokens": 64,
+  "min_tiles": 2,
+  "model_type": "lfm2_vl",
+  "pad_token_id": 0,
+  "projector_bias": true,
+  "projector_hidden_act": "gelu",
+  "projector_hidden_size": 2048,
+  "projector_use_layernorm": false,
+  "text_config": {
+    "_name_or_path": "LiquidAI/LFM2-350M",
+    "architectures": [
+      "Lfm2ForCausalLM"
+    ],
+    "block_auto_adjust_ff_dim": true,
+    "block_dim": 1024,
+    "block_ffn_dim_multiplier": 1.0,
+    "block_mlp_init_scale": 1.0,
+    "block_multiple_of": 256,
+    "block_norm_eps": 1e-05,
+    "block_out_init_scale": 1.0,
+    "block_use_swiglu": true,
+    "block_use_xavier_init": true,
+    "bos_token_id": 1,
+    "conv_L_cache": 3,
+    "conv_bias": false,
+    "conv_dim": 1024,
+    "conv_dim_out": 1024,
+    "conv_use_xavier_init": true,
+    "dtype": "bfloat16",
+    "eos_token_id": 7,
+    "full_attn_idxs": null,
+    "hidden_size": 1024,
+    "initializer_range": 0.02,
+    "intermediate_size": 6656,
+    "layer_types": [
+      "conv",
+      "conv",
+      "full_attention",
+      "conv",
+      "conv",
+      "full_attention",
+      "conv",
+      "conv",
+      "full_attention",
+      "conv",
+      "full_attention",
+      "conv",
+      "full_attention",
+      "conv",
+      "full_attention",
+      "conv"
+    ],
+    "max_position_embeddings": 128000,
+    "model_type": "lfm2",
+    "norm_eps": 1e-05,
+    "num_attention_heads": 16,
+    "num_heads": 16,
+    "num_hidden_layers": 16,
+    "num_key_value_heads": 8,
+    "pad_token_id": 0,
+    "rope_parameters": {
+      "rope_theta": 1000000.0,
+      "rope_type": "default"
+    },
+    "tie_word_embeddings": true,
+    "use_cache": true,
+    "use_pos_enc": true,
+    "vocab_size": 65536
+  },
+  "tie_word_embeddings": true,
+  "tile_size": 512,
+  "transformers_version": "5.8.1",
+  "use_image_special_tokens": true,
+  "use_thumbnail": true,
+  "vision_config": {
+    "attention_dropout": 0.0,
+    "dtype": "bfloat16",
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 768,
+    "intermediate_size": 3072,
+    "layer_norm_eps": 1e-06,
+    "model_type": "siglip2_vision_model",
+    "num_attention_heads": 12,
+    "num_channels": 3,
+    "num_hidden_layers": 12,
+    "num_patches": 256,
+    "patch_size": 16,
+    "vision_use_head": false
+  }
+}

generation_config.json

@@ -0,0 +1,10 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": [
+    7,
+    7
+  ],
+  "pad_token_id": 0,
+  "transformers_version": "5.8.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7e5694ca9a0ea81985c84822a503b3d583f4dc71fb2540d1bda944c57a16546
+size 897484568

model_eval/README.md

@@ -0,0 +1,203 @@
+# Eval pipeline (OpenRouter judge)
+
+A self-contained evaluation pipeline for LFM2.5-VL structured-extraction
+models. Extraction runs on your local GPU (vLLM/HF); the VLM judge runs
+remotely via the [OpenRouter](https://openrouter.ai/) API — no need to
+host a 30+ GB vision judge yourself.
+
+## Pipeline
+
+```
+WDS tars  ─▶  Extraction (local GPU)  ─▶  predictions
+                                              │
+              structural metrics  ◀───────────┤
+              (json validity, key P/R/F1)     │
+                                              │
+              VLM judge (OpenRouter)  ◀───────┘
+                            │
+                            ▼
+                       eval_result.json
+```
+
+Three primary metrics per run: `json_validity_rate`, `key_f1_macro`,
+`vlm_judge_score_avg` (per-key precision / recall also reported as
+diagnostic byproducts of F1).
+
+## Files
+
+```
+.
+├── README.md
+├── requirements.txt
+├── run_eval.sh       ← entry script (env vars + python call)
+├── run_eval.py       ← CLI + orchestrator + metrics aggregation
+├── extract.py        ← WDS loader + vLLM/HF extraction + JSON parsing
+├── judge.py          ← OpenRouter async VLM judging
+├── prompts/          ← 2 prompt templates (.txt)
+└── eval_data/        ← shipped 2000-sample eval set (single WDS tar)
+```
+
+Three Python files total. No nested packages, no `pyproject.toml`,
+no `pip install -e .` — just `pip install -r requirements.txt`.
+
+---
+
+## Setup
+
+### 1. Python environment
+
+```bash
+python -m venv .venv && source .venv/bin/activate
+pip install -r requirements.txt
+```
+
+`pip install` will pull `vllm`, `torch`, `transformers`, `peft`,
+`webdataset`, `pillow`, `openai`, `tqdm`, `numpy` — ~5 GB total, takes
+5–15 min depending on the network.
+
+> **Mac / no NVIDIA GPU?** vLLM won't install. Either drop the `vllm`
+> line from `requirements.txt`, or install everything else manually and
+> run with `--extraction-backend hf` (forces the HF transformers path).
+
+### 2. OpenRouter API key
+
+Get a key from https://openrouter.ai/keys, then add it to your `~/.bashrc`:
+
+```bash
+export OPENROUTER_API_KEY=sk-or-v1-...
+```
+
+Then `source ~/.bashrc` (or open a new shell).
+
+---
+
+## Run
+
+### Quick start
+
+```bash
+bash run_eval.sh
+```
+
+Defaults:
+- Evaluates `LiquidAI/LFM2.5-VL-450M-Extract` on `./eval_data/`
+- Runs the full **2000 samples** (~30 min)
+- VLM judge: `qwen/qwen3.5-35b-a3b`
+- Writes results to `./eval_result.json` and log to `./eval_run.log`
+
+### Tweaking knobs
+
+Open `run_eval.sh` — every knob is a top-level variable with an inline
+comment. Common changes:
+
+```bash
+NUM_SAMPLES=50                # set 50 for a quick smoke test (~5 min)
+EXTRACTION_BACKEND="hf"       # if vLLM init fails on your machine
+EXTRACTION_BATCH=32           # bump for faster extraction (default 8)
+VLM_JUDGE_MODEL="google/gemini-2.5-flash"   # any image-capable OpenRouter model id
+JUDGE_CONCURRENCY=8           # lower if you hit OpenRouter rate limits
+```
+
+### CLI alternative
+
+If you'd rather skip the .sh wrapper, drive `run_eval.py` directly:
+
+```bash
+python run_eval.py \
+  --checkpoint-path LiquidAI/LFM2.5-VL-450M-Extract \
+  --data-path ./eval_data \
+  --output-path ./eval_result.json \
+  --num-samples 50 \
+  --extraction-backend auto \
+  --vlm-judge --vlm-judge-model qwen/qwen3.5-35b-a3b
+```
+
+All flags: `python run_eval.py --help`
+
+---
+
+## Eval data
+
+### What ships in `./eval_data/`
+
+2000 `(image, schema, JSON)` samples in a single WebDataset tar
+(`eval_set_n2000.tar`). Reference labels were generated by an ensemble
+of frontier multimodal models and lightly post-processed for consistency.
+
+### Bring your own
+
+Drop a `.tar` (or directory of tars) anywhere and pass
+`--data-path /path/to/your/data`.
+
+### Format spec
+
+Each sample is a WebDataset group sharing a common `<sample_id>` prefix:
+
+```
+<sample_id>.jpg                image bytes
+<sample_id>.key_explanations   JSON {key_name: description}   (the schema)
+<sample_id>.structured_text    JSON {key_name: value}         (ground truth)
+```
+
+---
+
+## Output
+
+`./eval_result.json` has three top-level keys:
+
+```jsonc
+{
+  "metadata": {
+    "checkpoint_path": "LiquidAI/LFM2.5-VL-450M-Extract",
+    "num_samples_evaluated": 50,
+    "extraction_backend": "auto",
+    "vlm_judge_model": "qwen/qwen3.5-35b-a3b",
+    "elapsed_s": 215.2,
+    "timestamp_utc": "2026-05-29T..."
+  },
+  "metrics": {
+    "json_validity_rate":   0.996,    // share of samples with parseable JSON
+    "key_precision_macro":  0.996,    // pred-keys ∩ gt-keys / pred-keys
+    "key_recall_macro":     0.997,
+    "key_f1_macro":         0.997,    // primary schema-consistency metric
+    "vlm_judge_score_avg":  0.922,    // 0-1, VLM scoring of all keys vs image
+    "samples_evaluated":    50
+  },
+  "samples": [
+    /* per-sample {schema, gt, prediction, per_key scores, raw judge text} */
+  ]
+}
+```
+
+The `samples[].vlm_judge_raw` field preserves the judge's verbatim text
+response — useful for debugging unexpected scores.
+
+---
+
+## Costs
+
+Default judge on a full 2000-sample run, calculated against per-token
+pricing at the time of writing (check https://openrouter.ai/models for
+current rates):
+
+| Stage | Model | Input rate | Output rate | Est. cost |
+|---|---|---|---|---|
+| VLM judge | `qwen/qwen3.5-35b-a3b` | $0.139 / 1M | $1.00 / 1M | ~$1.53 |
+
+**Full 2000-sample run: ~$1.50.** Smoke 50-sample: ~$0.04.
+
+---
+
+## Troubleshooting
+
+- **vLLM init fails** (e.g. `Ninja build failed` / `__cudaLaunch not declared`)
+  → set `EXTRACTION_BACKEND="hf"` in `run_eval.sh` for a slower-but-stable
+  fallback.
+- **OpenRouter 429 (rate limit)** → lower `JUDGE_CONCURRENCY` to 4 or 8.
+- **`No usable samples loaded`** → your tars don't have the expected
+  `<key>.jpg` / `.key_explanations` / `.structured_text` fields, or the
+  `.tar` path is wrong.
+- **A new judge model rejects with `Reasoning is mandatory`** or returns all
+  zero scores with `finish_reason=length` → edit the `_VLM_JUDGE_REASONING`
+  constant in `judge.py` (the OpenRouter `reasoning` param works differently
+  per model).

model_eval/eval_data/eval_set_n2000.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:762bc2e15c5f2636a51ba6d985e8cec31bd0658a11e48a5f6e7e4bbbd5b71923
+size 135168000

model_eval/extract.py

@@ -0,0 +1,533 @@
+"""WDS data loading, schema-aware extraction prompts, local model inference,
+and JSON-from-noise parsing — everything the trained-checkpoint stage needs.
+
+Public entry: `run_extraction(samples, model_path, backend, ...)` returns a
+list of records ready for the judge stage.
+"""
+
+from __future__ import annotations
+
+import base64
+import io
+import json
+import logging
+import re
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from string import Template
+from typing import Any, Iterator, Literal
+
+import webdataset as wds
+
+logger = logging.getLogger(__name__)
+
+_IMAGE_EXTS = ("jpg", "jpeg", "png", "webp")
+_PROMPT_DIR = Path(__file__).resolve().parent / "prompts"
+_EXTRACTION_TPL = Template((_PROMPT_DIR / "extraction_system.txt").read_text(encoding="utf-8"))
+
+
+# ─── data loading ──────────────────────────────────────────────────────────
+
+
+@dataclass(frozen=True)
+class EvalSample:
+    key: str
+    image_bytes: bytes
+    schema: dict[str, str]
+    ground_truth: dict[str, object]
+
+
+def discover_tar_files(data_path: str) -> list[str]:
+    """Resolve a path/glob/brace-expansion to a sorted list of `.tar` files."""
+    if "{" in data_path and ".." in data_path:
+        expanded = list(wds.shardlists.expand_urls(data_path))
+        if expanded and Path(expanded[0]).is_dir():
+            tars: list[str] = []
+            for d in expanded:
+                if Path(d).is_dir():
+                    tars.extend(sorted(str(f) for f in Path(d).rglob("*.tar")))
+            if not tars:
+                raise FileNotFoundError(f"No .tar files found in: {data_path}")
+            return tars
+        return expanded
+
+    p = Path(data_path)
+    if p.is_file() and p.suffix == ".tar":
+        return [str(p)]
+    if p.is_dir():
+        tars = sorted(str(f) for f in p.rglob("*.tar"))
+        if not tars:
+            raise FileNotFoundError(f"No .tar files found in {data_path}")
+        return tars
+    parent = p.parent
+    tars = sorted(str(f) for f in parent.glob(p.name))
+    if not tars:
+        raise FileNotFoundError(f"No files matching pattern: {data_path}")
+    return tars
+
+
+def _first_image(sample: dict) -> bytes | None:
+    """Return the first image, preferring `imgN.jpg` order then legacy keys."""
+    multi: list[tuple[int, bytes]] = []
+    for k, v in sample.items():
+        if not isinstance(v, (bytes, bytearray)) or not k.startswith("img"):
+            continue
+        head, _, ext = k.partition(".")
+        if ext.lower() not in _IMAGE_EXTS:
+            continue
+        idx_str = head[3:]
+        if not idx_str.isdigit():
+            continue
+        multi.append((int(idx_str), bytes(v)))
+    if multi:
+        multi.sort(key=lambda x: x[0])
+        return multi[0][1]
+    for k in _IMAGE_EXTS:
+        v = sample.get(k)
+        if isinstance(v, (bytes, bytearray)):
+            return bytes(v)
+    return None
+
+
+def _decode_text(value: object) -> str:
+    if value is None:
+        return ""
+    if isinstance(value, bytes):
+        return value.decode("utf-8", errors="replace")
+    return str(value)
+
+
+def iter_eval_samples(
+    data_path: str,
+    *,
+    skip: int = 0,
+    limit: int = 0,
+) -> Iterator[EvalSample]:
+    """Yield up to `limit` EvalSamples from WDS tars.
+
+    Each sample carries `<key>.jpg`, `<key>.key_explanations` (schema with
+    descriptions), and `<key>.structured_text` (ground-truth values).
+    Samples missing image/schema/labels are silently skipped.
+    """
+    tar_files = discover_tar_files(data_path)
+    logger.info("Discovered %d tar file(s) under %s", len(tar_files), data_path)
+
+    dataset = wds.WebDataset(
+        tar_files,
+        shardshuffle=False,
+        nodesplitter=None,
+        handler=lambda e: logger.warning("WDS skip: %s", e) or True,
+    )
+    n_seen = 0
+    n_yielded = 0
+    for sample in dataset:
+        img = _first_image(sample)
+        ke = sample.get("key_explanations")
+        st = sample.get("structured_text")
+        if img is None or ke is None or st is None:
+            continue
+        try:
+            schema = json.loads(_decode_text(ke))
+            gt = json.loads(_decode_text(st))
+        except (json.JSONDecodeError, ValueError) as e:
+            logger.warning("Skip %s: bad JSON (%s)", sample.get("__key__", "?"), e)
+            continue
+        if not isinstance(schema, dict) or not isinstance(gt, dict):
+            continue
+        n_seen += 1
+        if n_seen <= skip:
+            continue
+        yield EvalSample(
+            key=str(sample.get("__key__", f"sample_{n_seen}")),
+            image_bytes=img,
+            schema=schema,
+            ground_truth=gt,
+        )
+        n_yielded += 1
+        if limit and n_yielded >= limit:
+            break
+    logger.info("Yielded %d eval sample(s) (skipped %d)", n_yielded, skip)
+
+
+# ─── prompt rendering ──────────────────────────────────────────────────────
+
+
+def schema_to_yaml(schema: dict[str, str]) -> str:
+    return "\n".join(f"{k}: {v}" for k, v in schema.items())
+
+
+def build_extraction_prompt(schema: dict[str, str]) -> str:
+    return _EXTRACTION_TPL.substitute(schema=schema_to_yaml(schema))
+
+
+# ─── JSON parsing ──────────────────────────────────────────────────────────
+
+
+def sanitize_output(text: str) -> str:
+    """Strip whitespace + markdown fences + bare `json` prefix."""
+    if not text:
+        return ""
+    s = text.strip()
+    if s.startswith("```"):
+        nl = s.find("\n")
+        s = "" if nl == -1 else s[nl + 1 :]
+        s = s.rstrip()
+        if s.endswith("```"):
+            s = s[:-3]
+        s = s.strip()
+    head = s.split("\n", 1)
+    if head and head[0].strip().lower() == "json":
+        s = head[1] if len(head) > 1 else ""
+        s = s.strip()
+    return s
+
+
+def _first_balanced(text: str, start: int) -> str | None:
+    """Return `text[start:i+1]` when braces balance; None if never balances."""
+    depth = 0
+    in_string = False
+    escape = False
+    for i in range(start, len(text)):
+        ch = text[i]
+        if escape:
+            escape = False
+            continue
+        if ch == "\\" and in_string:
+            escape = True
+            continue
+        if ch == '"':
+            in_string = not in_string
+            continue
+        if in_string:
+            continue
+        if ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return text[start : i + 1]
+    return None
+
+
+_TRAILING_COMMA_RE = re.compile(r",(\s*[}\]])")
+# Bare empty-string entries inside an object: ` "",` or `\n  ""\n}`.
+# Some VLMs emit these as a runaway-collapse pattern.
+_BARE_EMPTY_RE = re.compile(r',\s*""\s*(?=[,}])')
+_BARE_EMPTY_BEFORE_CLOSE_RE = re.compile(r',\s*""\s*(?=\n*\s*})')
+
+
+def extract_json_strict_first(text: str) -> tuple[dict, bool]:
+    """Sanitize + parse. Returns `(dict, was_strict)`.
+
+    `was_strict=True` if the strict parse succeeded — that's what
+    `json_valid` reports. False covers repaired-success and total failure
+    (caller distinguishes via `bool(dict)`).
+    """
+    sanitized = sanitize_output(text)
+    if not sanitized:
+        return {}, False
+    start = sanitized.find("{")
+    if start == -1:
+        return {}, False
+
+    candidate = _first_balanced(sanitized, start)
+    if candidate is not None:
+        try:
+            parsed = json.loads(candidate)
+            if isinstance(parsed, dict):
+                return parsed, True
+        except (json.JSONDecodeError, ValueError):
+            pass
+
+    # Second-chance repair (ported from old bundle's `_repair_parse`):
+    # try original `bal`, then progressively repaired versions, then the
+    # last-`}` truncation with both repairs applied. First dict wins.
+    candidates: list[str] = []
+    bal = _first_balanced(sanitized[start:], 0)
+    if bal is not None:
+        candidates.append(bal)
+        c2 = _BARE_EMPTY_RE.sub("", bal)
+        c2 = _BARE_EMPTY_BEFORE_CLOSE_RE.sub("", c2)
+        candidates.append(c2)
+        candidates.append(_TRAILING_COMMA_RE.sub(r"\1", c2))
+    last_close = sanitized.rfind("}")
+    if last_close >= 0:
+        tail = sanitized[: last_close + 1]
+        candidates.append(tail)
+        tail2 = _BARE_EMPTY_RE.sub("", tail)
+        tail2 = _BARE_EMPTY_BEFORE_CLOSE_RE.sub("", tail2)
+        tail2 = _TRAILING_COMMA_RE.sub(r"\1", tail2)
+        candidates.append(tail2)
+
+    for c in candidates:
+        try:
+            parsed = json.loads(c)
+        except (json.JSONDecodeError, ValueError):
+            continue
+        if isinstance(parsed, dict):
+            return parsed, False
+    return {}, False
+
+
+# ─── extraction backends ───────────────────────────────────────────────────
+
+
+def _img_to_data_url(img_bytes: bytes) -> str:
+    b64 = base64.b64encode(img_bytes).decode("ascii")
+    return f"data:image/jpeg;base64,{b64}"
+
+
+def _build_chat_messages(schema: dict[str, str], img_bytes: bytes) -> list[dict[str, Any]]:
+    return [
+        {"role": "system", "content": build_extraction_prompt(schema)},
+        {
+            "role": "user",
+            "content": [
+                {"type": "image_url", "image_url": {"url": _img_to_data_url(img_bytes)}},
+            ],
+        },
+    ]
+
+
+def _extract_vllm(
+    samples: list[EvalSample],
+    *,
+    model_path: str,
+    max_model_len: int,
+    gpu_mem_util: float,
+    max_new_tokens: int,
+) -> list[str]:
+    """vLLM offline batch extraction. One shot, no retries — Ctrl+C if hung."""
+    from vllm import LLM  # type: ignore
+
+    logger.info("Initializing vLLM for %s …", model_path)
+    llm = LLM(
+        model=model_path,
+        trust_remote_code=True,
+        dtype="bfloat16",
+        max_model_len=max_model_len,
+        gpu_memory_utilization=gpu_mem_util,
+        enable_prefix_caching=True,
+        disable_log_stats=True,
+        limit_mm_per_prompt={"image": 1},
+    )
+    from vllm import SamplingParams  # type: ignore
+
+    sp = SamplingParams(temperature=0.0, max_tokens=max_new_tokens)
+    conversations = [_build_chat_messages(s.schema, s.image_bytes) for s in samples]
+    logger.info("vLLM.chat over %d samples …", len(samples))
+    # Suppress reasoning for extraction-side reasoning models (Qwen3 family,
+    # gpt-oss family). Without this they burn the token budget on internal
+    # <think> blocks and emit no JSON. Non-reasoning models silently ignore.
+    outputs = llm.chat(
+        conversations,
+        sampling_params=sp,
+        use_tqdm=True,
+        chat_template_kwargs={
+            "enable_thinking": False,
+            "reasoning_effort": "low",
+        },
+    )
+    texts = [o.outputs[0].text if o.outputs else "" for o in outputs]
+    return texts
+
+
+def _extract_hf(
+    samples: list[EvalSample],
+    *,
+    model_path: str,
+    max_new_tokens: int,
+    batch: int,
+) -> list[str]:
+    """HF transformers fallback. Slower but works without vLLM (e.g. Mac)."""
+    import torch  # type: ignore
+    from PIL import Image  # type: ignore
+    from transformers import AutoModelForImageTextToText, AutoProcessor  # type: ignore
+
+    logger.info("Loading HF model %s …", model_path)
+    processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
+    # Decoder-only generation requires left padding so the model never sees
+    # padding tokens in the middle of the sequence at decode time.
+    if hasattr(processor, "tokenizer") and processor.tokenizer is not None:
+        processor.tokenizer.padding_side = "left"
+    model = AutoModelForImageTextToText.from_pretrained(
+        model_path,
+        dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
+        trust_remote_code=True,
+        device_map="auto" if torch.cuda.is_available() else None,
+    )
+    model.eval()
+
+    outputs: list[str] = []
+    for start in range(0, len(samples), batch):
+        chunk = samples[start : start + batch]
+        msgs = [_build_chat_messages(s.schema, s.image_bytes) for s in chunk]
+        # The processor strips the image_url data URIs and replaces with PIL.
+        for m, s in zip(msgs, chunk):
+            m[1]["content"][0] = {"type": "image", "image": Image.open(io.BytesIO(s.image_bytes))}
+        inputs = processor.apply_chat_template(
+            msgs,
+            add_generation_prompt=True,
+            tokenize=True,
+            return_dict=True,
+            return_tensors="pt",
+            padding=True,
+            # Suppress reasoning blocks (Qwen3 family) — kwarg flows into the
+            # model's Jinja chat template. Non-reasoning models ignore it.
+            enable_thinking=False,
+        ).to(model.device)
+        with torch.no_grad():
+            gen = model.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False)
+        decoded = processor.batch_decode(
+            gen[:, inputs["input_ids"].shape[1] :], skip_special_tokens=True
+        )
+        outputs.extend(decoded)
+        logger.info("HF extraction: %d/%d", min(start + batch, len(samples)), len(samples))
+    return outputs
+
+
+def _extract_smolvlm(
+    samples: list[EvalSample],
+    *,
+    model_path: str,
+    max_new_tokens: int,
+    max_model_len: int = 8192,
+    gpu_mem_util: float = 0.85,
+) -> list[str]:
+    """SmolVLM / Idefics3-family extraction via vLLM with user-prompt format.
+
+    Why a dedicated path:
+      - SmolVLM was trained on user/assistant turns only; system messages
+        carry weak signal and trigger generic image-captioning behavior
+        rather than schema following. So we put the schema in the *user*
+        prompt alongside the image.
+      - vLLM natively supports the Idefics3 architecture (SmolVLM v1/v2),
+        giving ~20× the throughput of single-sample HF generation. We use
+        it directly here instead of going through the generic vLLM path
+        (which would also work, but with a system-prompt template).
+    """
+    from vllm import LLM, SamplingParams  # type: ignore
+
+    logger.info("Initializing vLLM for SmolVLM/Idefics3 model: %s …", model_path)
+    llm = LLM(
+        model=model_path,
+        trust_remote_code=True,
+        dtype="bfloat16",
+        max_model_len=max_model_len,
+        gpu_memory_utilization=gpu_mem_util,
+        enable_prefix_caching=True,
+        disable_log_stats=True,
+        limit_mm_per_prompt={"image": 1},
+    )
+    sp = SamplingParams(temperature=0.0, max_tokens=max_new_tokens)
+
+    conversations: list[list[dict[str, Any]]] = []
+    for s in samples:
+        b64 = base64.b64encode(s.image_bytes).decode("ascii")
+        data_url = f"data:image/jpeg;base64,{b64}"
+        # User prompt (no system) — schema goes in the user turn alongside
+        # the image. This is the format SmolVLM responds to.
+        conversations.append([
+            {"role": "user", "content": [
+                {"type": "image_url", "image_url": {"url": data_url}},
+                {"type": "text", "text": build_extraction_prompt(s.schema)},
+            ]},
+        ])
+
+    logger.info("vLLM.chat over %d samples (SmolVLM) …", len(samples))
+    outputs = llm.chat(conversations, sampling_params=sp, use_tqdm=True)
+    return [o.outputs[0].text if o.outputs else "" for o in outputs]
+
+
+def _is_smolvlm(model_path: str) -> bool:
+    """Detect SmolVLM / Idefics3-family models from path."""
+    p = model_path.lower()
+    return "smolvlm" in p or "idefics" in p
+
+
+def run_extraction(
+    samples: list[EvalSample],
+    *,
+    model_path: str,
+    backend: Literal["auto", "vllm", "hf"] = "auto",
+    max_new_tokens: int = 1024,
+    max_model_len: int = 8192,
+    gpu_mem_util: float = 0.85,
+    batch: int = 8,
+) -> list[dict[str, Any]]:
+    """Run extraction; return one prediction record per input sample.
+
+    `backend="auto"` tries vLLM first and falls back to HF on import error
+    or init failure. `"vllm"` / `"hf"` force the choice.
+
+    Special case: SmolVLM / Idefics3 family always uses a dedicated code
+    path regardless of `backend` — vLLM doesn't support them well, and the
+    standard `AutoModelForImageTextToText` invocation drops the chat
+    template specifics they need.
+    """
+    if not samples:
+        return []
+
+    t0 = time.perf_counter()
+
+    # SmolVLM / Idefics: dedicated path, bypass `backend` selection.
+    if _is_smolvlm(model_path):
+        logger.info("Detected SmolVLM/Idefics-family model — using dedicated extraction path.")
+        raw_outputs = _extract_smolvlm(samples, model_path=model_path, max_new_tokens=max_new_tokens)
+        backend_used = "smolvlm"
+    elif backend == "hf":
+        raw_outputs = _extract_hf(samples, model_path=model_path, max_new_tokens=max_new_tokens, batch=batch)
+        backend_used = "hf"
+    elif backend == "vllm":
+        raw_outputs = _extract_vllm(
+            samples,
+            model_path=model_path,
+            max_model_len=max_model_len,
+            gpu_mem_util=gpu_mem_util,
+            max_new_tokens=max_new_tokens,
+        )
+        backend_used = "vllm"
+    else:  # auto
+        try:
+            raw_outputs = _extract_vllm(
+                samples,
+                model_path=model_path,
+                max_model_len=max_model_len,
+                gpu_mem_util=gpu_mem_util,
+                max_new_tokens=max_new_tokens,
+            )
+            backend_used = "vllm"
+        except Exception as e:
+            logger.warning("vLLM extraction failed (%s); falling back to HF transformers.", e)
+            raw_outputs = _extract_hf(samples, model_path=model_path, max_new_tokens=max_new_tokens, batch=batch)
+            backend_used = "hf"
+
+    dt = time.perf_counter() - t0
+    logger.info(
+        "Extraction over %d samples took %.1fs (%.2f sample/s, backend=%s).",
+        len(samples),
+        dt,
+        len(samples) / max(dt, 1e-9),
+        backend_used,
+    )
+
+    if len(raw_outputs) != len(samples):
+        raise RuntimeError(
+            f"Backend returned {len(raw_outputs)} outputs for {len(samples)} samples"
+        )
+
+    records: list[dict[str, Any]] = []
+    for s, raw in zip(samples, raw_outputs):
+        parsed, strict = extract_json_strict_first(raw)
+        records.append(
+            {
+                "key": s.key,
+                "schema": s.schema,
+                "ground_truth": s.ground_truth,
+                "prediction_raw": raw,
+                "prediction_json": parsed,
+                "prediction_strict_valid": strict,
+            }
+        )
+    return records

model_eval/judge.py

@@ -0,0 +1,296 @@
+"""OpenRouter-backed VLM judge.
+
+One prompt per sample (batched-over-keys in one JSON answer), run
+concurrently via asyncio. The VLM judge scores every key against the
+image and is the only quality signal we ship.
+
+Requires `OPENROUTER_API_KEY` in the environment. OpenRouter is
+OpenAI-compatible, so we point the `openai` SDK's `base_url` at it.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import base64
+import json
+import logging
+import os
+import re
+from pathlib import Path
+from string import Template
+from typing import Any
+
+from openai import AsyncOpenAI
+from tqdm.asyncio import tqdm_asyncio
+
+logger = logging.getLogger(__name__)
+
+_PROMPT_DIR = Path(__file__).resolve().parent / "prompts"
+_VLM_JUDGE_TPL = Template((_PROMPT_DIR / "vlm_judge_batch.txt").read_text(encoding="utf-8"))
+
+OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1"
+
+# Reasoning suppression for the VLM judge. Default model
+# (`qwen/qwen3.5-35b-a3b`) requires `enable_thinking=False` to avoid
+# burning the full token budget on internal thought and returning
+# `finish_reason=length` with no visible output. Both layers (top-level
+# `reasoning` + `chat_template_kwargs`) are sent so whichever the
+# underlying provider honors gets used; the other is ignored.
+_VLM_JUDGE_REASONING = {
+    "reasoning": {"enabled": False},
+    "chat_template_kwargs": {"enable_thinking": False},
+}
+
+
+# ─── per-key pre-classification ────────────────────────────────────────────
+
+
+def normalize_bool(gt_val: object, pred_val: object) -> object:
+    """If GT is bool and pred is a yes/no/true/false string, coerce."""
+    if isinstance(gt_val, bool) and isinstance(pred_val, str):
+        lower = pred_val.strip().lower()
+        if lower in ("yes", "true"):
+            return True
+        if lower in ("no", "false"):
+            return False
+    return pred_val
+
+
+def initialize_per_key_evals(records: list[dict[str, Any]]) -> None:
+    """Build a `per_key` mapping `{key: {gt, pred}}` for the VLM judge to score."""
+    for rec in records:
+        per_key: dict[str, dict[str, Any]] = {}
+        gt = rec["ground_truth"]
+        pred = rec["prediction_json"]
+        for key, gt_val in gt.items():
+            pred_val = pred.get(key)
+            if pred_val is not None:
+                pred_val = normalize_bool(gt_val, pred_val)
+            per_key[key] = {"gt": gt_val, "pred": pred_val}
+        rec["per_key"] = per_key
+
+
+# ─── prompt building ───────────────────────────────────────────────────────
+
+
+def _vlm_attr_block(entries: list[tuple[str, str, object]]) -> str:
+    lines: list[str] = []
+    for key, desc, pred in entries:
+        lines.append(f'- "{key}": {desc}\n  predicted: {json.dumps(pred, ensure_ascii=False)}')
+    return "\n".join(lines)
+
+
+def build_vlm_judge_prompt(entries: list[tuple[str, str, object]]) -> str:
+    return _VLM_JUDGE_TPL.substitute(attributes=_vlm_attr_block(entries))
+
+
+# ─── response parsing ──────────────────────────────────────────────────────
+
+
+_NUM_RE = re.compile(r"(\d+\.?\d*)")
+
+
+def _clamp(x: float) -> float:
+    return 0.0 if x < 0.0 else (1.0 if x > 1.0 else x)
+
+
+def parse_score(text: str | None) -> float:
+    if not text:
+        return 0.0
+    text = text.strip()
+    try:
+        return _clamp(float(text))
+    except ValueError:
+        m = _NUM_RE.search(text)
+        if not m:
+            return 0.0
+        try:
+            return _clamp(float(m.group(1)))
+        except ValueError:
+            return 0.0
+
+
+def parse_batch_scores(text: str | None, expected_keys: list[str]) -> dict[str, float]:
+    """Parse `{key: score}` from judge output; missing keys default to 0.0."""
+    result = {k: 0.0 for k in expected_keys}
+    if not text:
+        return result
+    # Reuse the JSON extractor from extract.py — same logic.
+    from extract import extract_json_strict_first
+
+    parsed, _ = extract_json_strict_first(text)
+    if not isinstance(parsed, dict):
+        return result
+    for k in expected_keys:
+        v = parsed.get(k)
+        if isinstance(v, bool):
+            # Defensive: bool is a subclass of int. Don't accept it as a score.
+            continue
+        if isinstance(v, (int, float)):
+            result[k] = _clamp(float(v))
+        elif isinstance(v, str):
+            # Judge sometimes returns string-formatted numbers like "0.8";
+            # parse_score extracts the leading numeric.
+            result[k] = parse_score(v)
+    return result
+
+
+def per_sample_judge_avg(per_key: dict[str, dict[str, Any]], score_field: str) -> float | None:
+    """Average a judge score across all keys of one sample. None if no scores."""
+    scores = [
+        float(v.get(score_field))
+        for v in per_key.values()
+        if isinstance(v.get(score_field), (int, float)) and not isinstance(v.get(score_field), bool)
+    ]
+    return sum(scores) / len(scores) if scores else None
+
+
+# ─── OpenRouter client + concurrent dispatch ───────────────────────────────
+
+
+def _img_to_data_url(img_bytes: bytes) -> str:
+    return f"data:image/jpeg;base64,{base64.b64encode(img_bytes).decode('ascii')}"
+
+
+def _make_client(api_key: str | None) -> AsyncOpenAI:
+    key = api_key or os.environ.get("OPENROUTER_API_KEY")
+    if not key:
+        raise RuntimeError(
+            "OPENROUTER_API_KEY is not set. Get a key at https://openrouter.ai/keys "
+            "and `export OPENROUTER_API_KEY=...` before running."
+        )
+    return AsyncOpenAI(base_url=OPENROUTER_BASE_URL, api_key=key, max_retries=3, timeout=120.0)
+
+
+async def _one_chat(
+    client: AsyncOpenAI,
+    *,
+    model: str,
+    system: str,
+    user_text: str,
+    image_bytes: bytes | None,
+    max_tokens: int,
+    semaphore: asyncio.Semaphore,
+    extra_body: dict[str, Any] | None = None,
+    max_retries_on_empty: int = 3,
+) -> str:
+    """Single OpenRouter chat call, rate-limited by `semaphore`.
+
+    Some OpenRouter providers occasionally return HTTP 200 with empty content
+    (no model output). Treating that as success silently fails the per-key
+    JSON parse — so we retry up to `max_retries_on_empty` times before
+    giving up.
+    """
+    user_content: list[dict[str, Any]] = [{"type": "text", "text": user_text}]
+    if image_bytes:
+        user_content.insert(0, {"type": "image_url", "image_url": {"url": _img_to_data_url(image_bytes)}})
+    messages = [
+        {"role": "system", "content": system},
+        {"role": "user", "content": user_content},
+    ]
+    kwargs: dict[str, Any] = {
+        "model": model,
+        "messages": messages,
+        "max_tokens": max_tokens,
+        "temperature": 0.0,
+    }
+    if extra_body:
+        kwargs["extra_body"] = extra_body
+
+    async with semaphore:
+        for attempt in range(max_retries_on_empty + 1):
+            try:
+                resp = await client.chat.completions.create(**kwargs)
+                text = resp.choices[0].message.content or ""
+                if text.strip():
+                    return text
+                finish = getattr(resp.choices[0], "finish_reason", None)
+                if attempt < max_retries_on_empty:
+                    logger.warning(
+                        "Empty response from %s (finish_reason=%s); retrying %d/%d.",
+                        model, finish, attempt + 1, max_retries_on_empty,
+                    )
+                    continue
+                logger.warning(
+                    "Empty response from %s after %d retries (finish_reason=%s); giving up.",
+                    model, max_retries_on_empty, finish,
+                )
+                return ""
+            except Exception as e:
+                logger.warning("OpenRouter call failed (%s); returning empty string.", e)
+                return ""
+        return ""
+
+
+async def _run_concurrent(
+    coros: list[Any],
+    *,
+    desc: str,
+) -> list[str]:
+    """Run coroutines concurrently with a tqdm progress bar."""
+    return await tqdm_asyncio.gather(*coros, desc=desc)
+
+
+# ─── VLM judge ─────────────────────────────────────────────────────────────
+
+
+def run_vlm_judge(
+    records: list[dict[str, Any]],
+    *,
+    sample_images: dict[str, bytes],
+    model: str,
+    max_tokens: int = 1024,
+    concurrency: int = 16,
+    api_key: str | None = None,
+) -> None:
+    """Score every key against the image via an OpenRouter VLM."""
+    plans: list[dict[str, Any]] = []
+    prompts: list[str] = []
+    imgs: list[bytes] = []
+    for rec in records:
+        per_key = rec.get("per_key", {})
+        if not per_key:
+            continue
+        keys = list(per_key.keys())
+        entries = [(k, rec["schema"].get(k, ""), per_key[k].get("pred")) for k in keys]
+        img = sample_images.get(rec["key"], b"")
+        if not img:
+            logger.warning("VLM judge: missing image for sample %s", rec["key"])
+            continue
+        prompts.append(build_vlm_judge_prompt(entries))
+        imgs.append(img)
+        plans.append({"record": rec, "keys": keys})
+
+    if not plans:
+        for rec in records:
+            rec["vlm_judge_avg"] = None
+        return
+
+    logger.info("VLM judge: scoring %d sample(s) × all keys via %s.", len(plans), model)
+
+    client = _make_client(api_key)
+    sem = asyncio.Semaphore(concurrency)
+    coros = [
+        _one_chat(
+            client,
+            model=model,
+            system="You are a meticulous visual evaluator.",
+            user_text=p,
+            image_bytes=img,
+            max_tokens=max_tokens,
+            semaphore=sem,
+            extra_body=_VLM_JUDGE_REASONING,
+        )
+        for p, img in zip(prompts, imgs)
+    ]
+    raw_outputs = asyncio.run(_run_concurrent(coros, desc="VLM judge"))
+
+    for plan, text in zip(plans, raw_outputs):
+        scores = parse_batch_scores(text, plan["keys"])
+        rec = plan["record"]
+        rec["vlm_judge_raw"] = text
+        for k in plan["keys"]:
+            rec["per_key"][k]["vlm_score"] = scores.get(k, 0.0)
+
+    for rec in records:
+        rec["vlm_judge_avg"] = per_sample_judge_avg(rec["per_key"], "vlm_score")

model_eval/prompts/extraction_system.txt

@@ -0,0 +1,5 @@
+Extract the following from the image:
+
+$schema
+
+Respond with only a JSON object. Do not include any text outside the JSON.

model_eval/prompts/vlm_judge_batch.txt

@@ -0,0 +1,17 @@
+Look at the image and evaluate whether the predicted values are correct for each attribute.
+
+Attributes (with descriptions and predictions):
+$attributes
+
+Each prediction should be an observation about the image (a word, number, or natural-language phrase like "partly cloudy" or "lying horizontally"), NOT a restatement of the description/question itself.
+
+For each attribute, rate how correct the prediction is based on what you see in the image, on a 0.0-1.0 scale:
+- 1.0: Exactly correct or semantically identical
+- 0.7-0.9: Mostly correct with minor differences
+- 0.3-0.6: Partially correct
+- 0.0-0.2: Incorrect, unrelated, or just rephrases the description instead of giving a specific value
+
+Respond with ONLY a JSON object mapping each attribute name to its score. Example:
+{"attr1": 0.9, "attr2": 0.3, "attr3": 1.0}
+
+No text outside the JSON.

model_eval/requirements.txt

@@ -0,0 +1,16 @@
+# --- Extraction (local) ---
+vllm>=0.20.1                 # optional but recommended on NVIDIA GPUs; pass
+                             # --extraction-backend hf to skip if it won't install.
+transformers>=4.45
+torch>=2.4
+peft>=0.10
+webdataset>=0.2
+pillow>=10.0
+num2words>=0.5               # SmolVLM/SmolVLM2 processor requires this for video timestamp formatting
+
+# --- OpenRouter judges (remote) ---
+openai>=1.50                 # OpenRouter is OpenAI-compatible; just point base_url at it.
+
+# --- Tiny helpers ---
+tqdm>=4.66
+numpy>=1.24

model_eval/run_eval.py

@@ -0,0 +1,209 @@
+#!/usr/bin/env python
+"""Eval pipeline driver: extraction → structural metrics → VLM judge → JSON.
+
+Extraction runs locally on your GPU (vLLM/HF); the VLM judge runs remotely
+via the OpenRouter API. One process, sequential stages, one JSON file out.
+"""
+
+from __future__ import annotations
+
+import argparse
+import datetime as _dt
+import json
+import logging
+import sys
+import time
+from pathlib import Path
+from typing import Any
+
+from extract import iter_eval_samples, run_extraction
+from judge import initialize_per_key_evals, run_vlm_judge
+
+
+# ─── metrics aggregation ───────────────────────────────────────────────────
+
+
+def per_sample_structural(prediction_json: dict, ground_truth: dict, strict_valid: bool) -> dict[str, Any]:
+    pred_keys = set(prediction_json.keys())
+    gt_keys = set(ground_truth.keys())
+    overlap = pred_keys & gt_keys
+    p = len(overlap) / len(pred_keys) if pred_keys else 0.0
+    r = len(overlap) / len(gt_keys) if gt_keys else 0.0
+    f1 = 2 * p * r / (p + r) if (p + r) > 0 else 0.0
+    return {
+        "json_valid": strict_valid,
+        "total_keys": len(gt_keys),
+        "total_pred_keys": len(pred_keys),
+        "overlap_keys": len(overlap),
+        "key_precision": p,
+        "key_recall": r,
+        "key_f1": f1,
+    }
+
+
+def aggregate(records: list[dict[str, Any]]) -> dict[str, Any]:
+    n = len(records)
+    if n == 0:
+        return {"samples_evaluated": 0}
+
+    def mean(xs: list[float]) -> float:
+        return sum(xs) / len(xs) if xs else 0.0
+
+    json_valid = sum(1 for r in records if r.get("json_valid"))
+    vlm_scores = [r["vlm_judge_avg"] for r in records if r.get("vlm_judge_avg") is not None]
+
+    return {
+        "json_validity_rate": json_valid / n,
+        "key_precision_macro": mean([r.get("key_precision", 0.0) for r in records]),
+        "key_recall_macro": mean([r.get("key_recall", 0.0) for r in records]),
+        "key_f1_macro": mean([r.get("key_f1", 0.0) for r in records]),
+        "vlm_judge_score_avg": mean(vlm_scores) if vlm_scores else None,
+        "samples_evaluated": n,
+    }
+
+
+def _strip_sample(rec: dict[str, Any]) -> dict[str, Any]:
+    """Drop heavy/internal fields before serialising to JSON."""
+    return {
+        "key": rec["key"],
+        "schema": rec["schema"],
+        "ground_truth": rec["ground_truth"],
+        "prediction_raw": rec["prediction_raw"],
+        "prediction_json": rec["prediction_json"],
+        "json_valid": rec.get("json_valid", False),
+        "total_keys": rec.get("total_keys", 0),
+        "total_pred_keys": rec.get("total_pred_keys", 0),
+        "key_precision": rec.get("key_precision", 0.0),
+        "key_recall": rec.get("key_recall", 0.0),
+        "key_f1": rec.get("key_f1", 0.0),
+        "vlm_judge_avg": rec.get("vlm_judge_avg"),
+        "vlm_judge_raw": rec.get("vlm_judge_raw"),
+        "per_key": rec.get("per_key", {}),
+    }
+
+
+# ─── CLI ───────────────────────────────────────────────────────────────────
+
+
+def main() -> int:
+    p = argparse.ArgumentParser(description="OpenRouter-judged structured-extraction eval.")
+    p.add_argument("--checkpoint-path", required=True, help="HF id or local merged/LoRA dir.")
+    p.add_argument("--data-path", default="./eval_data", help="WDS tar / dir / glob.")
+    p.add_argument("--output-path", default="./eval_result.json")
+    p.add_argument("--num-samples", type=int, default=0, help="Cap N samples (0 = all).")
+    p.add_argument("--skip-samples", type=int, default=0)
+    p.add_argument("--extraction-backend", choices=["auto", "vllm", "hf"], default="auto")
+    p.add_argument("--extraction-batch", type=int, default=8)
+    p.add_argument("--extraction-max-new-tokens", type=int, default=1024)
+    p.add_argument("--extraction-gpu-mem-util", type=float, default=0.85)
+    p.add_argument("--extraction-max-model-len", type=int, default=8192)
+    p.add_argument("--vlm-judge", action=argparse.BooleanOptionalAction, default=True)
+    p.add_argument("--vlm-judge-model", default="qwen/qwen3-vl-4b-instruct")
+    p.add_argument("--vlm-judge-max-tokens", type=int, default=1024)
+    p.add_argument("--judge-concurrency", type=int, default=16, help="Concurrent OpenRouter calls.")
+    p.add_argument("--openrouter-api-key", default=None, help="Override $OPENROUTER_API_KEY.")
+    p.add_argument("--log-level", default="INFO")
+    args = p.parse_args()
+
+    logging.basicConfig(
+        level=args.log_level.upper(),
+        format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+    )
+
+    t_start = time.perf_counter()
+    logger = logging.getLogger("run_eval")
+    logger.info("=== OpenRouter-judged eval starting ===")
+
+    # ── load samples ─────────────────────────────────��───────────────────
+    samples = list(
+        iter_eval_samples(
+            args.data_path,
+            skip=args.skip_samples,
+            limit=args.num_samples,
+        )
+    )
+    if not samples:
+        raise RuntimeError(
+            f"No usable samples loaded from {args.data_path} — expected WDS tars "
+            "with .jpg, .key_explanations, .structured_text per sample."
+        )
+    logger.info("Loaded %d sample(s).", len(samples))
+    sample_images = {s.key: s.image_bytes for s in samples}
+
+    # ── extraction ───────────────────────────────────────────────────────
+    records = run_extraction(
+        samples,
+        model_path=args.checkpoint_path,
+        backend=args.extraction_backend,
+        max_new_tokens=args.extraction_max_new_tokens,
+        max_model_len=args.extraction_max_model_len,
+        gpu_mem_util=args.extraction_gpu_mem_util,
+        batch=args.extraction_batch,
+    )
+
+    # ── structural metrics ───────────────────────────────────────────────
+    for rec in records:
+        rec.update(
+            per_sample_structural(
+                rec["prediction_json"],
+                rec["ground_truth"],
+                rec.get("prediction_strict_valid", bool(rec["prediction_json"])),
+            )
+        )
+
+    initialize_per_key_evals(records)
+    judge_errors: dict[str, str] = {}
+
+    # ── VLM judge ────────────────────────────────────────────────────────
+    if args.vlm_judge:
+        try:
+            run_vlm_judge(
+                records,
+                sample_images=sample_images,
+                model=args.vlm_judge_model,
+                max_tokens=args.vlm_judge_max_tokens,
+                concurrency=args.judge_concurrency,
+                api_key=args.openrouter_api_key,
+            )
+        except Exception as e:
+            judge_errors["vlm_judge"] = repr(e)
+            logger.warning("VLM judge failed (%s); continuing without VLM scores.", e)
+            for rec in records:
+                rec.setdefault("vlm_judge_avg", None)
+    else:
+        for rec in records:
+            rec["vlm_judge_avg"] = None
+
+    # ── write output ─────────────────────────────────────────────────────
+    elapsed = time.perf_counter() - t_start
+    result = {
+        "metadata": {
+            "checkpoint_path": args.checkpoint_path,
+            "data_path": args.data_path,
+            "num_samples_evaluated": len(records),
+            "extraction_backend": args.extraction_backend,
+            "vlm_judge_model": args.vlm_judge_model if args.vlm_judge else None,
+            "judge_errors": judge_errors or None,
+            "elapsed_s": round(elapsed, 2),
+            "timestamp_utc": _dt.datetime.now(_dt.timezone.utc).isoformat(),
+        },
+        "metrics": aggregate(records),
+        "samples": [_strip_sample(rec) for rec in records],
+    }
+
+    out = Path(args.output_path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+    out.write_text(json.dumps(result, ensure_ascii=False, indent=2), encoding="utf-8")
+
+    print()
+    print("=== JUDGING SUMMARY ===")
+    print(f"output={out}")
+    for k, v in result["metrics"].items():
+        print(f"  {k}={v:.4f}" if isinstance(v, float) else f"  {k}={v}")
+    print(f"  elapsed_s={elapsed:.1f}")
+    print("=== JUDGING OK ===")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

model_eval/run_eval.sh

@@ -0,0 +1,67 @@
+#!/bin/bash
+# Example evaluation script — extraction on local GPU, judges via OpenRouter API.
+#   bash run_eval.sh
+
+set -euo pipefail
+
+# --- vLLM env workarounds ----------------------------------------------------
+# Needed when EXTRACTION_BACKEND=vllm or "auto". No-op on systems without
+# vLLM / environment-modules.
+#   1. CUDA toolkit on LD_LIBRARY_PATH so flashinfer's GDN/Mamba kernels can
+#      dlopen libcudart.so.12 (LFM2.5-VL is a hybrid arch and uses these).
+#   2. Skip flashinfer's sampling kernel — flashinfer 0.6 + CUDA 12.9 trigger
+#      an NVCC stub bug (`__cudaLaunch` not declared). PyTorch-native sampler
+#      is ~20% slower but works.
+#   3. Skip vLLM 0.21's DeepGEMM autotune warmup (~18 min for MoE/FP8 models).
+module load cuda12.9/toolkit/12.9.1 2>/dev/null || true
+export VLLM_USE_FLASHINFER_SAMPLER="${VLLM_USE_FLASHINFER_SAMPLER:-0}"
+export VLLM_USE_DEEP_GEMM="${VLLM_USE_DEEP_GEMM:-0}"
+
+# --- OpenRouter API key ------------------------------------------------------
+# Required. Get one at https://openrouter.ai/keys, then either:
+#   export OPENROUTER_API_KEY=...
+# in your shell, OR uncomment and set it here.
+#OPENROUTER_API_KEY="sk-or-v1-..."
+
+# --- Checkpoint --------------------------------------------------------------
+# HF id of the trained model, OR a local merged/LoRA checkpoint dir.
+CHECKPOINT="LiquidAI/LFM2.5-VL-450M-Extract"
+
+# --- Eval data ---------------------------------------------------------------
+# WDS tar / dir of tars / brace-glob.
+DATA_PATH="./eval_data"
+
+# Output JSON path.
+OUTPUT="./eval_result.json"
+
+# --- Sample count ------------------------------------------------------------
+# Number of samples to evaluate. Default 2000 runs the full shipped eval_data
+# (~30 min). Set to 50 for a quick smoke test (~5 min).
+NUM_SAMPLES=2000
+
+# --- Extraction (local GPU) --------------------------------------------------
+# "auto" tries vLLM first, falls back to HF transformers on init failure.
+EXTRACTION_BACKEND="auto"
+EXTRACTION_BATCH=8
+
+# --- Judge model (OpenRouter) ------------------------------------------------
+# Any image-capable OpenRouter model id works. Pricing:
+# https://openrouter.ai/models
+VLM_JUDGE_MODEL="qwen/qwen3.5-35b-a3b"
+
+# Concurrent OpenRouter calls. Lower if you hit rate limits.
+JUDGE_CONCURRENCY=16
+
+# --- Run ---------------------------------------------------------------------
+LOG_FILE="${LOG_FILE:-./eval_run.log}"
+echo "Logging to: ${LOG_FILE}"
+python run_eval.py \
+  --checkpoint-path "${CHECKPOINT}" \
+  --data-path "${DATA_PATH}" \
+  --output-path "${OUTPUT}" \
+  --num-samples "${NUM_SAMPLES}" \
+  --extraction-backend "${EXTRACTION_BACKEND}" \
+  --extraction-batch "${EXTRACTION_BATCH}" \
+  --vlm-judge --vlm-judge-model "${VLM_JUDGE_MODEL}" \
+  --judge-concurrency "${JUDGE_CONCURRENCY}" \
+  2>&1 | tee "${LOG_FILE}"

processor_config.json

@@ -0,0 +1,39 @@
+{
+  "image_processor": {
+    "data_format": "channels_first",
+    "do_image_splitting": true,
+    "do_normalize": true,
+    "do_pad": true,
+    "do_rescale": true,
+    "do_resize": true,
+    "downsample_factor": 2,
+    "encoder_patch_size": 16,
+    "image_mean": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "image_processor_type": "Lfm2VlImageProcessor",
+    "image_std": [
+      0.5,
+      0.5,
+      0.5
+    ],
+    "max_image_tokens": 256,
+    "max_num_patches": 1024,
+    "max_pixels_tolerance": 2.0,
+    "max_tiles": 10,
+    "min_image_tokens": 64,
+    "min_tiles": 2,
+    "resample": 2,
+    "rescale_factor": 0.00392156862745098,
+    "return_row_col_info": true,
+    "size": {
+      "height": 512,
+      "width": 512
+    },
+    "tile_size": 512,
+    "use_thumbnail": true
+  },
+  "processor_class": "Lfm2VlProcessor"
+}

tokenizer_config.json

@@ -0,0 +1,33 @@
+{
+  "backend": "tokenizers",
+  "bos_token": "<|startoftext|>",
+  "clean_up_tokenization_spaces": true,
+  "do_image_splitting": true,
+  "eos_token": "<|im_end|>",
+  "extra_special_tokens": [],
+  "image_end_token": "<|image_end|>",
+  "image_start_token": "<|image_start|>",
+  "image_thumbnail": "<|img_thumbnail|>",
+  "image_token": "<image>",
+  "is_local": true,
+  "legacy": false,
+  "local_files_only": false,
+  "max_length": null,
+  "model_max_length": 1000000000000000019884624838656,
+  "model_specific_special_tokens": {
+    "image_end_token": "<|image_end|>",
+    "image_start_token": "<|image_start|>",
+    "image_token": "<image>"
+  },
+  "pad_to_multiple_of": null,
+  "pad_token": "<|pad|>",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "processor_class": "Lfm2VlProcessor",
+  "return_token_type_ids": false,
+  "sp_model_kwargs": {},
+  "spaces_between_special_tokens": false,
+  "tokenizer_class": "TokenizersBackend",
+  "use_default_system_prompt": false,
+  "use_fast": true
+}