weather-intent / README.md
Nicholas55555's picture
Upload folder using huggingface_hub
8845b7f verified
|
Raw
History Blame Contribute Delete
2.12 kB
metadata
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - intent-parsing
  - structured-output
  - synthetic
  - weather
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: eval
        path: eval.jsonl

weather-intent

Synthetic dataset for fine-tuning a small model to parse a natural-language weather question into a compact structured intent (JSON), scored by field-level exact match. Trains Nicholas55555/qwen2.5-1.5b-weather-intent.

{"utterance": "will it rain in Paris this weekend?",
 "intent": {"location": "Paris", "date": "this_weekend", "metric": "rain", "time_of_day": null},
 "category": "basic"}

Fields

field description
utterance the natural-language weather question
intent target slots: location, date, metric, time_of_day
category failure-mode bucket (below)

Slot vocabulary

slot values
location place name, or null
date today, tomorrow, day_after_tomorrow, this_weekend, next_weekend, next_week, a weekday, or null
metric temperature, rain, snow, wind, humidity, uv, cloud, general
time_of_day morning, afternoon, evening, night, or null

Splits

split rows
train 1200
eval 217

The eval split is category-balanced with a hand-authored hard set. category buckets each example by failure mode: basic, multi_slot, implicit_slot, date_reasoning, vague, ood.

Generation

Programmatic and correct-by-construction: slots are sampled first, the utterance is rendered from them, so labels are exact. Deliberate phrasing variety (prefixes, slot ordering, synonymous stems); an optional distillation mode paraphrases seeds with a frontier model. The task is parsing only — volatile forecast facts and any rendering live in downstream code, never in the labels.