README.md

---
library_name: peft
base_model: LiquidAI/LFM2.5-VL-450M
license: cc-by-nc-4.0
tags:
- peft
- lora
- trl
- sft
- lfm2.5-vl
- satellite-imagery
- paired-image
- civilian-infrastructure
- conflict-disruption
- blackline-atlas
datasets:
- ChrisRPL/satellite-civilian-conflict-disruption-reporter-v1
---

# LFM2.5-VL Civilian Conflict Disruption Reporter LoRA v1

Diagnostic PEFT LoRA adapter for [LiquidAI/LFM2.5-VL-450M](https://huggingface.co/LiquidAI/LFM2.5-VL-450M), trained on [ChrisRPL/satellite-civilian-conflict-disruption-reporter-v1](https://huggingface.co/datasets/ChrisRPL/satellite-civilian-conflict-disruption-reporter-v1).

## Status

This is a **diagnostic local LoRA artifact**, not the accepted Blackline Atlas adapter and not the demo-critical runtime model.

Important audit note: the attempted Hugging Face Jobs run `69f22721d2c8bd8662bd3151` failed before training because the job command referenced a local absolute path that is not visible inside the remote HF Jobs container. Therefore this repo should not be described as a completed HF Jobs training result.

The current canonical Blackline adapter artifact remains `ChrisRPL/blackline-atlas-lfm25-vl-sft-train-hf-aux-v10-adapter`, and that adapter is also published only as a rejected research artifact after failing the eval-gold schema/action smoke gate.

## Intended task

Paired baseline/current satellite-image reporting for macro-visible civilian disruption from conflict, bombardment, explosion, shelling, or related human-caused disruption. The output is evidence-first JSON, not tactical guidance.

The adapter must **not** be used for military asset detection, route intelligence, targeting, strike support, or ranking of targets.

## Expected output schema

```json
{
  "visible_change_summary": "string",
  "civilian_disruption_evidence": ["collapsed_building", "debris_field"],
  "negative_evidence": ["no_visible_change"],
  "uncertainty_factors": ["string"],
  "severity_hint": "none | low | medium | high",
  "recommended_action": "discard | defer | downlink_now",
  "confidence": 0.0,
  "short_rationale": "string"
}
```

## Training summary

- Method: TRL `SFTTrainer` VLM SFT with PEFT LoRA.
- Base model: `LiquidAI/LFM2.5-VL-450M`.
- Dataset: `ChrisRPL/satellite-civilian-conflict-disruption-reporter-v1`.
- Diagnostic subset: 32 train rows, 16 eval rows.
- Epochs: 1.
- Batch size: 1, gradient accumulation: 8.
- Learning rate: `5e-5`, cosine scheduler, warmup ratio `0.05`.
- LoRA: `r=8`, `alpha=16`, `dropout=0.05`, `target_modules=all-linear`.
- Local run final train loss: `8.816`; eval loss: `7.937`.
- Trackio dashboard: https://huggingface.co/spaces/ChrisRPL/mlintern-lfm25v1

This is a diagnostic adapter, not a production model.

## Promotion status

Not promoted. Before any future promotion, rerun this training path from a self-contained HF Jobs script or bundle, then require eval-gold generation to produce schema-valid JSON, improved action match, nonzero `downlink_now` recall, and no false-positive regression.

## Known limitations

- Very small diagnostic run; loss remains high.
- Labels are partly inherited/rule-derived, not all expert-reviewed.
- BRIGHT rows are optical-to-SAR and introduce cross-modality artifacts.
- Local generation evaluation showed the base model did not satisfy the strict schema; adapter evaluation is limited and should be repeated on GPU before demo promotion.
- Associated HF Job `69f22721d2c8bd8662bd3151` failed before training and is not evidence of successful remote fine-tuning.