NuTonic/lspace

NU:TONIC satellite VLM — supervised fine-tuned (SFT) checkpoint derived from LiquidAI/LFM2.5-VL-450M on a single LEAP vlm_sft run over one mixed Parquet corpus (main + repeated task hubs + repeated Firewatch).

• Model page: https://huggingface.co/NuTonic/lspace
• Training recipe https://github.com/josephrp/nutonic : NU:TONIC — train/run_sat_vl_sft_e2e.py orchestrates train/materialize_vlm_sft_mix.py → LEAP vlm_sft via train/train_lfm_vl_sft.py and refs/leap-finetune-main.

Intended use

Use this model when you want a small (~0.45B) image–text model that has seen many supervised examples of:

• Satellite RGB chips (Sentinel-2–style optical previews / tiled chips used in NU:TONIC datasets),
• Optional overhead / map-style context stills (mapbox_stills/ in the training corpora),
• Optional analysis-condition visuals (profile-conditioned render PNGs present in some training rows),
• Multi-image user turns (temporal pairs and terramind predictions),
• Assistant outputs that mix narrative geospatial reasoning with structured artifacts seen in training, including normalized bounding boxes and JSON-like detection lists when prompted.

Typical applications:

• Satellite image captioning and coarse land-cover / structure description (non-exhaustive).
• Scenario-aligned narratives consistent with NU:TONIC “PRO mini-app” training slices:
  • wildfire / burn scar style reasoning (Firewatch-SFT slice),
  • coastal / bright-target / maritime-style reasoning (OceanScout-SFT slice),
  • land-cover transition reasoning (LandShift-SFT slice),
  • inundation / water-expansion reasoning (FloodPulse-SFT slice),
  • structured analytical brief writing (BriefComposer-SFT slice).

This checkpoint is not a full analytic pipeline: it does not fetch imagery from STAC, run Earth Engine, or guarantee calibration to real-world hazard operations without human review.

Training data (what it actually saw)

Training is main-heavy by construction: the mix streams almost all rows from the aggregate Hub dataset, then upsamples smaller hubs so rare behaviors still receive gradient mass after global shuffling.

Main corpus (dominant mass)

• NuTonic/sat-vl-sft-training-ready-v1
  Aggregate training-ready Parquet packaging NU:TONIC satellite VLM supervision derived from multiple builders, including (non-exhaustively) metadata-first procedural rows and bounding-box-heavy corpora. Rows commonly include messages with multi-part user.content mixing image + text, and assistant targets describing imagery, evidence, and/or structured outputs consistent with NU:TONIC JSONL/VLM conventions.

Upsampled task hubs (default repeat = 8× each)

These teach multi-image / vertical-specific behaviors described in internal NU:TONIC dataset planning (PRO mini-apps alignment):

• NuTonic/brief-composer-sft-v1 — mixed multi-image prompts toward structured analytical brief writing.
• NuTonic/oceanscout-sft-v1 — maritime / water-context bbox + narrative patterns.
• NuTonic/floodpulse-sft-v1 — temporal pair reasoning around inundation extent patterns.
• NuTonic/landshift-sft-v1 — temporal pair reasoning around land-cover transition patterns.

Upsampled small hub (default repeat = 80×)

• NuTonic/firewatch-sft-v1 — wildfire / burn scar oriented supervision (small row count; repeated for mass).

Important implication

Because SFT matches teacher strings, the model may:

• Echo dataset-specific prompt framing (profile cues, task wording),
• Prefer bbox conventions seen in training (typically 0–1 normalized box coordinates embedded in assistant text / JSON-like structures; see NU:TONIC notes aligned with LEAP vlm_sft conventions),
• Reflect English supervision dominate if that is true in the upstream datasets.

Non-goals / limitations

• No warranty of geophysical correctness: outputs are learned correlations from curated supervision; validate operationally for your AOI, sensor, season, and labeling definition.
• Distribution shift: performance drops are expected off-domain (different sensors, resolutions, projections, stylizations, heavy cloud cover, night imagery, SAR, etc.).
• Privacy / safety: training mixes may include overhead context stills in some rows; do not use outputs as sole evidence for high-risk decisions (disasters, enforcement, insurance) without independent verification.
• Grounding reliability: bbox/JSON outputs should be treated as model proposals, not GIS truth.

Inference quickstart (Transformers)

This family loads like other HF multimodal chat models (requires trust_remote_code=True for Liquid remote modules).

Minimal pattern (single image) — (AutoModelForImageTextToText + AutoProcessor):

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

model_id = "NuTonic/lspace"

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

pil = Image.open("chip.png").convert("RGB")
user_text = (
    "The input is satellite imagery (RGB). Describe surface cover and structure where visible, "
    "and note uncertainty."
)

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": pil},
            {"type": "text", "text": user_text},
        ],
    }
]

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():
    out = model.generate(**inputs, max_new_tokens=512, do_sample=False)

# Trim prompt tokens (exact slicing depends on model wrapper); simplest decode:
text = processor.batch_decode(out, skip_special_tokens=True)[0]
print(text)

# NuTonic/lspace

Fine-tuned from `LiquidAI/LFM2.5-VL-450M` using the NU:TONIC satellite VLM SFT mix
(`train/run_sat_vl_sft_e2e.py`): single LEAP run on main + task + Firewatch Parquet mix.

Training stack: LEAP `vlm_sft` in this repo's `refs/leap-finetune-main`.