Unsprawl-4B-SFT

Fine-tuned LoRA adapter for the Unsprawl platform's local inference pipeline.

Training

• Base model: Qwen3-4B-Instruct (4-bit quantized via Unsloth)
• Method: QLoRA (r=16, alpha=32, 0.81% trainable params)
• Data: 195 curated pairs (entity extraction, strategy generation, quality judging, tone classification)
• Training: 1 epoch, batch size 8, AdamW 8-bit, lr=2e-4
• Final loss: 1.69 (from 2.65 starting)
• Hardware: NVIDIA RTX 3060 12GB, 164 seconds

Tasks

Task	Description
Entity extraction	Structured JSON entity extraction from legal documents
Quality judging	5-dimension scoring with APPROVE/REJECT verdicts
Tone classification	6-category tone classification (NEUTRAL, INFLAMMATORY, etc.)
Claim verification	Factual claim extraction for KG cross-reference

Usage

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Mansib/Unsprawl-4B-SFT",
    max_seq_length=4096,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

Part of Unsprawl

Unsprawl is a domain-agnostic, mission-driven compound AI platform for autonomous infrastructure resilience. This adapter enables local inference for structured tasks, reducing API dependency on cloud LLMs.