Skip to content

STATUS.md

Latest commit

 

History

History
125 lines (103 loc) · 5.51 KB

STATUS.md

File metadata and controls

125 lines (103 loc) · 5.51 KB

Project Status

Last updated: 2026-05-25

Dataset Overview

Metric Value
Frequency range 280 - 1100 MHz
Countries FR, US, UK, CN, DE, RU, ES, IT, CH (9)
Total entries 500
Unique services/applications 178
Q&A training pairs 1500
Fact-check corrections applied 16
Sources consulted for verification 11 regulatory bodies

Pipeline Phases

Phase 1 -- Data Collection & Enrichment DONE

  • Baseline frequency tables (280-1100 MHz, 9 countries)
  • Fine-grained segment split (5 sub-bands: 280-400, 400-470, 470-700, 700-870, 870-1100)
  • Per-country allocations with real device examples
  • Modulation types, channel spacing, power limits
  • Regulatory references (ITU, FCC, ANFR, Ofcom, BNetzA, MIIT, CNAF, Roskomnadzor, BAKOM)
  • Merged dataset: merged_dataset/enriched_280_1100_mhz_ALL.json (500 entries)
  • CSV export: merged_dataset/enriched_280_1100_mhz_ALL.csv

Phase 2 -- Fact-Checking DONE

  • 5 independent fact-check reports per sub-band
  • Consolidated report: factcheck_reports/factcheck_CONSOLIDATED_280_1100_mhz.json
  • 16 critical corrections applied (see below)
  • Cross-referenced against official regulatory databases

Key Corrections Applied

# Issue Fix
1 315 MHz TPMS listed for EU EU uses 433.92 MHz, 315 MHz is US/JP only
2 TETRA listed for US/CN/RU US=P25 (NTIA), CN=PDT 350-370, RU=limited
3 WMTS 608-614 listed for EU US-only (FCC Part 95H)
4 LoRaWAN 868 listed for US/CN US=902-928 MHz, CN=470-510 MHz
5 UK 700 MHz Vodafone Corrected to O2/VMO2 (Ofcom 2021 auction)
6 FR 700 MHz Free/Bouygues swap ARCEP Dec 2015: Free=2x5, Bouygues=2x10
7 ECC Decision (14)02 cited Corrected to (15)01 for 700 MHz duplex gap
8 Gazpar gas meters at 868 MHz Corrected to 169 MHz VHF Wize (GRDF)
9 AEHF listed as UHF MILSATCOM Corrected to MUOS (AEHF is EHF 44 GHz)
10 COSPAS-SARSAT at 399.9 MHz Corrected to 406.0-406.1 MHz
11 BeiDou/GLONASS at 399.9 MHz Both are L-band only, removed
12 DME Y-mode 30 us Corrected to 36 us (ICAO Annex 10)
13 ITU 5.328A for ADS-B Corrected to 5.328B
14 863-870 labeled ISM Corrected to SRD (ETSI EN 300 220)
15 CH listed as TETRA CH uses TETRAPOL (Polycom network)
16 CN listed TETRA 380 MHz CN uses PDT 350-370 MHz

Phase 3 -- Q&A Dataset Generation DONE

  • 1500 Q&A pairs generated
  • 3 categories: fonction, reglementation, appareils_utilisateurs
  • Source citations in every answer
  • Format: JSONL (HuggingFace-ready) + JSON
  • Output: qa_dataset/spectrum_qa_dataset.jsonl

Phase 4 -- Validation Pipeline READY (not yet run)

8-step automated pipeline in Data_Process/:

Step Script Backend Status
01 Ingestion multi-sources git + datasets Code ready
02 Dedup (hash + semantic) sentence-transformers cuda:0 Code ready
03 RF validation + multi-source cross-check + confidence % rules + HF + Wikipedia + Web/forums Code ready
04 Protocol DB matching rapidfuzz CPU Code ready
05 LLM hallucination check transformers + device_map="auto" (Qwen 32B sharded across 6 GPUs) Code ready
06 NLI fact verification DeBERTa-v3 cuda:1 Code ready
07 Scoring (0-13 scale) rules Code ready
08 Export (JSONL buckets + confidence_pct) rules Code ready

Dual scoring system:

  1. Step 07 score (0-13) — internal quality checks (freq/protocol/timings/LLM/NLI/dedup)
  2. Step 03 confidence (0-100%) — external triangulation across HF + Wikipedia + WebSearch + forums, with bonuses for triangulation and penalties for contradictions / missing sources

Target rig: 6x RTX 3070 (48 GB VRAM), Ubuntu 24.04

Pending:

  • Transfer scripts to rig
  • Install vLLM (pip install vllm) and download Qwen/Qwen2.5-32B-Instruct HF weights
  • Run full pipeline end-to-end (with CROSS_CHECK=1)
  • Human audit of 'partial' bucket via audit_dashboard.py
  • Merge audit decisions back into final dataset

Phase 5 -- HuggingFace Publication TODO

  • Upload verified.jsonl as HF dataset
  • Dataset card with methodology
  • License selection

File Structure

Flipper_Zero_RF_DataSet/
  baseline/                      Original CSV frequency tables
  enriched_data/                 5 enriched JSON files per sub-band
  merged_dataset/                Merged 500-entry JSON + CSV
  qa_dataset/                    1500 Q&A pairs (JSONL + JSON)
  factcheck_reports/             5 sub-band reports + consolidated
  Data_Process/
    README.md                    Pipeline architecture docs
    scripts/
      00_config.py               Global config & backend dispatch
      01_ingest_sources.py       Multi-source ingestion
      02_dedup.py                Hash + semantic deduplication
      03_rf_validation.py        RF rules validation
      04_protocols_db.py         Protocol DB matching (rapidfuzz)
      05_llm_hallucination.py    LLM hallucination detection (Ollama)
      06_fact_verification.py    NLI fact-check (DeBERTa)
      07_scoring.py              Score aggregation & bucketing
      08_export.py               Final JSONL export + manifest
      audit_dashboard.py         Human audit CLI
      run_pipeline.sh            Pipeline orchestrator
  STATUS.md                      This file
  STANDARDS.md                   RF standards & regulatory references
  README.md                      Project overview