Vector-LFM2.5-VL-450M

<center> <div style="display: flex; justify-content: center; gap: 0.5em;"> <a href="https://huggingface.co/Alfaxad/Vector-LFM2.5-VL-450M"><strong>Model</strong></a> | <a href="https://huggingface.co/datasets/Alfaxad/vector-100k"><strong>Dataset</strong></a> | <a href="https://huggingface.co/Alfaxad/Vector-LFM2.5-VL-450M/tree/main/vectoros_run_artifacts"><strong>Run Artifacts</strong></a> | </div> </center> <br>

Vector-LFM2.5-VL-450M

Vector-LFM2.5-VL-450M is a variant of LiquidAI/LFM2.5-VL-450M fine-runed for geospatial epidemiology tasks.

The model was supervised on Alfaxad/vector-100k, a 100k-example multimodal dataset built from geospatial image packets, aligned environmental layers, vector/disease weak labels, and task prompts. It is intended to produce population-level environmental risk summaries, structured JSON outputs, source provenance, uncertainty audits, and field-task briefs from geospatial tiles.

Vector-LFM2.5-VL-450M is not a diagnostic model. It does not confirm disease presence, assign individual risk, or replace field epidemiology, clinical judgment, or public-health authority review.

Model Details

Vector-LFM2.5-VL-450M inherits the general vision-language architecture and chat template of LFM2.5-VL-450M, then adapts it to the VectorOS remote-sensing task distribution.

• Base model: LiquidAI/LFM2.5-VL-450M
• Fine-tuned model: Alfaxad/Vector-LFM2.5-VL-450M
• Dataset: Alfaxad/vector-100k
• Fine-tuning method: supervised fine-tuning with LoRA, then merged into a full model
• LoRA configuration: rank 16, alpha 32, dropout 0.05
• Training examples: 72,000
• Validation examples: 14,000
• Test examples: 14,000
• Training length limit: 3,584 tokens
• Fixed re-evaluation length limit: 4,096 tokens
• Fixed re-evaluation generation limit: 1,600 new tokens
• Maximum image tokens used during training/evaluation: 256
• Training hardware: NVIDIA H100 80GB HBM3 on Modal
• Training runtime: 9,698.18 seconds
• Final train loss: 0.12335

Inherited LFM2.5-VL properties from the base model include the LFM2.5 language backbone, SigLIP2-based vision encoder, multimodal chat template, and support for image-text-to-text generation through Hugging Face Transformers.

VectorOS Task Scope

The fine-tune teaches the model to follow VectorOS-specific geospatial instructions over image/map packets and aligned metadata. The model is trained to reason over:

• Sentinel-2 and Mapbox-derived visual context
• rainfall signals
• surface-water context
• land-cover context
• population exposure
• OSM and health-facility proximity
• weak vector/disease evidence labels
• uncertainty and source provenance

The supervised task families include:

Chat Template

Use the LFM2.5-VL multimodal chat template through processor.apply_chat_template. A typical VectorOS prompt combines one image packet with a concise instruction.

<|startoftext|><|im_start|>system
You are VectorOS, a geospatial public-health assistant. Provide population-level surveillance support only.<|im_end|>
<|im_start|>user
<image>Analyze this geospatial tile and return a risk_tile_json output for the configured AOI and disease module.<|im_end|>
<|im_start|>assistant
...

Inference

Install current Transformers support for LFM2.5-VL:

pip install "transformers>=5.1.0" pillow accelerate safetensors

Run image-text-to-text inference:

from PIL import Image
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "Alfaxad/Vector-LFM2.5-VL-450M"

model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    device_map="auto",
    dtype=torch.bfloat16,
)
processor = AutoProcessor.from_pretrained(model_id)

image = Image.open("vectoros_tile_packet.png").convert("RGB")

conversation = [
    {
        "role": "system",
        "content": [
            {
                "type": "text",
                "text": (
                    "You are VectorOS, a geospatial public-health assistant. "
                    "Provide population-level surveillance support only."
                ),
            }
        ],
    },
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {
                "type": "text",
                "text": (
                    "Analyze this geospatial tile and return risk_tile_json. "
                    "Use only evidence visible in the tile packet and provided metadata. "
                    "Do not claim field-confirmed disease presence."
                ),
            },
        ],
    },
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    tokenize=True,
).to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,
    )

new_tokens = outputs[:, inputs["input_ids"].shape[-1]:]
answer = processor.batch_decode(new_tokens, skip_special_tokens=True)[0]
print(answer)

For long structured outputs such as full provenance or uncertainty audits, increase max_new_tokens up to 1600, which was the fixed evaluation setting.

Fine-Tuning

The model was trained for one epoch using LoRA SFT and then merged into a full model repository. The adapter and checkpoint metadata are retained under vectoros_run_artifacts/.

Training output:

Evaluation

• loads the base model from LiquidAI/LFM2.5-VL-450M
• loads the merged model directly from Alfaxad/Vector-LFM2.5-VL-450M
• evaluates loss on the full 14k validation split and full 14k test split
• evaluates generation on a 300-record stratified test sample
• uses max_new_tokens=1600 to avoid clipping long JSON outputs
• fixes JSON extraction, source-grounding recall, and negated safety-pattern handling

Artifacts:

• vectoros_run_artifacts/eval_comparison_fixed.json
• vectoros_run_artifacts/reeval_fixed_hf/
• vectoros_run_artifacts/train_config.json
• vectoros_run_artifacts/train_output.json

Held-Out Loss

Generation Evaluation

Generation evaluation used 300 stratified test records.

By-task median sequence similarity:

Interpretation

The fine-tuned model is substantially better than the base model on the geospatial epidemiology task distribution. The largest gains are in held-out likelihood, format following, source recall, structured output validity, and task-specific language. Risk-class accuracy improved from 0% to 43.33% on the 300-record generation sample, but this remains the most important metric to improve before using the model as a primary risk classifier.

Artifact Layout

This repository stores the merged model at the repo root and training/evaluation artifacts under vectoros_run_artifacts/.

Safety Scope

Vector-LFM2.5-VL-450M should only be used for geospatial epidemiology support, education, research, and VectorOS product prototyping.

Do not use this model for:

• individual diagnosis
• individual risk assignment
• clinical triage or treatment guidance
• claims of field-confirmed disease presence
• autonomous public-health interventions
• emergency response decisions without expert review

The model is trained on weak labels and geospatial proxies. Outputs should be treated as decision-support intelligence requiring source review and local public-health validation.

Limitations

• Evaluation is strongest for the Vector dataset/task distribution and does not prove field-level epidemiological validity.
• The model may overfit VectorOS phrasing because the supervised dataset has consistent schemas and task templates.
• Risk class accuracy is improved but still limited on generated risk_tile_json samples.
• The model does not access live geospatial feeds. It only reasons from the supplied image packet and prompt context.
• Public-health deployment requires review against local policy, disease-module assumptions, and data provenance.

Contact

This model was built for the VectorOS hackathon project by Alfaxad. Use the Hugging Face repository discussions for model-specific questions and artifact issues.

Citation

@misc{vectorlfm25vl,
  title = {Vector-LFM2.5-VL-450M: A Vision-Language Model for Geospatial Epidemiology},
  author = {Alfaxad},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Alfaxad/Vector-LFM2.5-VL-450M}}
}

@article{liquidai2025lfm2,
  title = {LFM2 Technical Report},
  author = {Liquid AI},
  journal = {arXiv preprint arXiv:2511.23404},
  year = {2025}
}