Instructions to use LiquidAI/LFM2.5-VL-450M-Extract with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use LiquidAI/LFM2.5-VL-450M-Extract with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="LiquidAI/LFM2.5-VL-450M-Extract")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("LiquidAI/LFM2.5-VL-450M-Extract")
model = AutoModelForImageTextToText.from_pretrained("LiquidAI/LFM2.5-VL-450M-Extract")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use LiquidAI/LFM2.5-VL-450M-Extract with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "LiquidAI/LFM2.5-VL-450M-Extract"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LiquidAI/LFM2.5-VL-450M-Extract",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/LiquidAI/LFM2.5-VL-450M-Extract

SGLang

How to use LiquidAI/LFM2.5-VL-450M-Extract with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "LiquidAI/LFM2.5-VL-450M-Extract" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LiquidAI/LFM2.5-VL-450M-Extract",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "LiquidAI/LFM2.5-VL-450M-Extract" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "LiquidAI/LFM2.5-VL-450M-Extract",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use LiquidAI/LFM2.5-VL-450M-Extract with Docker Model Runner:
```
docker model run hf.co/LiquidAI/LFM2.5-VL-450M-Extract
```

LFM2.5-VL-450M-Extract / model_eval /README.md

zhkleciel

Super-squash branch 'main' using huggingface_hub

c013d55 3 days ago

preview code

raw

history blame contribute delete

6.15 kB

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 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

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:

export OPENROUTER_API_KEY=sk-or-v1-...

Then source ~/.bashrc (or open a new shell).

Run

Quick start

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:

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:

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:

{
  "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).