Casual Chat Utterance Data Collector

Project Overview We are building a GR13 topic classifier and need high-quality training samples for the casual/social chat class. You will source, write, or curate standalone casual-chat utterances drawn from real user interaction patterns and deliver them in JSONL format. This is a paid contract. Compensation is per validated batch of 500 samples. Shortlisted candidates will be invited to a small paid trial before the full contract. No unpaid work is required at any stage, including during the application process. What You Will Do Source casual/social utterances from publicly available, real-user interaction datasets (Reddit casual conversation, Twitter/X small-talk corpora, open chat logs, etc.) Validate each sample against the inclusion criteria — standalone, casual tone, no implicit context required Clean and normalise text (fix encoding issues, remove usernames / PII) Format each accepted sample as a single JSONL record and submit as a .jsonl file Provide a short sourcing note per batch (dataset name, URL, filtering steps applied) Inclusion and Exclusion Criteria Include: Greetings and small talk Expressions of mood or emotion General opinions and casual questions with no domain dependency Jokes, friendly remarks, and everyday conversational exchanges Exclude: Customer-service or help requests, technical questions, product/service queries Messages that only make sense with prior conversational context Multi-turn dialogues — one standalone utterance per record only Anything belonging to another GR13 class (booking, finance, health, legal, etc.) Any message requiring a domain label to interpret its intent Required JSONL Format Each line must follow this exact schema. Each record contains five fields: id -- Sequential identifier, format: cc_0001, cc_0002, etc. utterance -- The cleaned, standalone casual message label -- Always: casual_chat source -- Dataset name or origin identifier (e.g. reddit_casual, twitter_smalltalk) lang -- Language code (en for English) Example records (formatted for readability): id: "cc_0001" utterance: "Haha yeah same, I have no idea what I'm doing today" label: "casual_chat" source: "reddit_casual" lang: "en" id: "cc_0002" utterance: "Good morning! Hope your week is going well" label: "casual_chat" source: "twitter_smalltalk" lang: "en" The actual .jsonl file must be one valid JSON object per line. Deliverables Per Batch JSONL Sample File — 500+ validated utterances per batch Label Accuracy — 95% or above (spot-check reviewed) Sourcing Log — .txt or .csv documenting dataset name, URL, and filtering steps applied Skills and Requirements NLP or text annotation experience (required) Data sourcing from public corpora Comfortable working with JSONL / JSON formats English — native or C2 level Familiarity with casual internet language and slang Strong attention to labelling edge cases How to Apply Please submit a proposal that includes the following. No unpaid sample files are required — we evaluate candidates on experience and proposed approach. A brief description of your NLP or text annotation experience The public data sources you plan to use (dataset names and URLs) Your estimated turnaround time for 500 validated samples A link to any relevant past work or portfolio (if available)