freight/README.md

Freight & Logistics Dispatch AI Lab (LoadETA)

This directory contains domain-specific research, synthetic dataset generation engines, and fine-tuning recipes for autonomous freight negotiation voice agents.

Directory Layout

freight/
├── data/
│   ├── freight_negotiation_omniroute.jsonl  # High-reasoning synthetic dataset generated via OmniRoute 'paid-premium'
│   └── freight_negotiation_sample.jsonl     # Procedural multi-turn SFT dataset with OpenAI tool-calling schema
├── docs/
│   └── FREIGHT_INTELLIGENCE_REPORT.md       # 51k-character deep domain report on US freight econometrics, broker tactics, accessorials, and FMCSA rules
└── scripts/
    ├── generate_freight_dataset.py          # Procedural seed-matrix generator with LoadETA tool schemas
    ├── omniroute_freight_synthesizer.py     # OmniRoute 'paid-premium' multi-agent dialogue synthesizer
    └── train_qwen_lora.py                   # 4-bit QLoRA training recipe for Qwen 3.8 9B with GGUF export instructions

Quick Start

1. Synthesize Training Data via OmniRoute

python3 freight/scripts/omniroute_freight_synthesizer.py --count 50 --output freight/data/freight_negotiation_omniroute.jsonl

2. Fine-Tune Qwen 3.8 9B

python3 freight/scripts/train_qwen_lora.py \
  --model_id empero-ai/Qwen3.8-9B \
  --dataset_path freight/data/freight_negotiation_omniroute.jsonl \
  --epochs 3

3. Deploy to Space / LiveKit

Convert the fused weights to GGUF (Q4_K_M) and mount under /data in abalanescu/flow ZeroGPU Space or host directly in your homelab.