daredevil467/hanoi-weather-router-data-v7

Hanoi Weather Router — Training Data v7 (Multi-turn ChatML)

Multi-task intent + scope + rewrite classification data for Hanoi weather chatbot router SLM. v7 introduces multi-turn ChatML format with sliding window K=3.

Files

File	Rows	Purpose
multitask_train_v7.jsonl	3470	v7 training — multi-turn ChatML schema
multitask_val_v7.jsonl	380	v7 validation — stratified split
system_prompt_v7.txt	1	Router system prompt v7 (1.7k chars)

v7 Changes (vs v6)

Architecture: Multi-turn ChatML messages format

Replaced single-shot context: {location, intent, turn} dict với multi-turn ChatML messages: [{role, content}] format (sliding window K=3, router-only history).

Research-backed (CANARD/QReCC/CONQRR/CHIQ): proven format cho conversational query rewriting. Generalizes anaphora resolution (location, time, intent) without per-issue field invention.

Data quality fixes

• 357 manual FIX samples: prefix collision cores (Cầu Giấy → Phường default), POI banned per P8/P9 policy, time anchor preservation.
• 250 multi-turn samples: turn=2/3 anaphora resolution (location/time inheritance, explicit-switch patterns).
• 100 pure-ward samples: full coverage 126/126 wards + 30/30 districts.
• 80 collision counter-examples: T4 borderline tier + explicit-switch boost.
• 346 POI samples DROPPED: hard-removed per P8/P9 commit (no silent fallback).

Confidence: 5-tier calibration scheme

Tier	Conf	Tình huống
T1	0.92	Direct query đầy đủ entity + intent + time
T2	0.85	Standard inference (1 default needed)
T3	0.80	Multi-turn inherited anchor
T4	0.74	Collision core không prefix
T5	0.62	Smalltalk / POI / OOD (rewrite=null)

Aligns with per-intent thresholds 0.73-0.85 → enables real calibration ECE measurement.

Schema

Each line is a JSON object:

{
  "history": [
    {"user": "Phường Cầu Giấy thứ 2 tuần sau thế nào?",
     "assistant": "{\"intent\":\"daily_forecast\",\"scope\":\"ward\",\"confidence\":0.92,\"rewritten_query\":\"Dự báo Phường Cầu Giấy thứ 2 tuần sau thế nào?\"}"}
  ],
  "input": "thứ 3 thì sao?",
  "output": {
    "intent": "daily_forecast",
    "scope": "ward",
    "confidence": 0.80,
    "rewritten_query": "Dự báo Phường Cầu Giấy thứ 3 tuần sau thế nào?"
  }
}

• history: array of {user, assistant} pairs, max K=3 (sliding window). Empty [] cho turn 1.
• assistant content: 4-keys core JSON (intent/scope/confidence/rewritten_query), JSON-stringified.
• input: current user query.
• output: 4-keys core JSON.

Training Prompt format (ChatML multi-turn)

<|im_start|>system
{SYSTEM_PROMPT}
<|im_end|>
[for each (user, assistant) in history:]
  <|im_start|>user
  {user}
  <|im_end|>
  <|im_start|>assistant
  {assistant}
  <|im_end|>
<|im_start|>user
{input}
<|im_end|>
<|im_start|>assistant
{output_json}
<|im_end|>

Intents (15 classes)

current_weather, hourly_forecast, daily_forecast, weather_overview, rain_query, temperature_query, wind_query, humidity_fog_query, historical_weather, location_comparison, activity_weather, expert_weather_param, weather_alert, seasonal_context, smalltalk_weather

Scopes (3 classes)

city, district, ward

Distribution

• Turn distribution: 75% turn=1, 24% turn=2, ~1% turn=3
• Scope: 31% city, 26% district, 44% ward
• Tier: T1=22%, T2=27%, T3=22%, T4=22%, T5=7%
• Coverage: 126/126 wards (100%), 30/30 districts (100%) đều mention trong manual data