Skip to content

data_spec.md

Latest commit

 

History

History
40 lines (34 loc) · 2.46 KB

data_spec.md

File metadata and controls

40 lines (34 loc) · 2.46 KB
Raw

data/classifications_for_analysis.csv — File Spec

Overview

Label-level classification data for conversational programming sessions. Each row represents one (message, label) pair. Multi-label messages produce multiple rows.

Granularity

  • Unit of analysis: one label assigned to one message
  • Message: identified by (sha, index_in_chat)
  • Session: identified by sha
  • Repository: identified by repository_full_name

Columns

Column Type Description
repository_full_name str GitHub repo identifier, e.g. owner/repo
sha str Session ID (unique per Cursor/Copilot chat session)
timestamp datetime (UTC) Session-level timestamp — all messages in a session share the same value. Use for chronological ordering of sessions within a repo (cross-session analysis).
title str Auto-generated session title
index_in_chat int 0-based message position within session
index_in_chat_original int Original 0-based message position within session (before filtering out "8.1 Others" messages)
truncated_content str Truncated user message text (used for classification context)
labels str (JSON) Raw LLM output — list of label objects with reasoning, main_category, sub_category
content str Full user message text
label_rank int 1-based rank of this label within the message (1 = dominant label)
reasoning str Flattened: LLM reasoning for this specific label
main_category str Flattened: e.g. "2. Failure Reporting"
sub_category str Flattened: e.g. "2.2 Symptom Description"

Key Properties

  • Multi-label: a message with N labels appears as N rows with label_rank 1..N
  • Single-label message: exactly one row with label_rank = 1
  • Message-level weight: when aggregating, use label_weight = 1 / count(rows per (sha, index_in_chat)) to avoid over-counting multi-label messages
  • Session ordering: sort by (repository_full_name, timestamp, sha) for cross-session analysis; sort by (sha, index_in_chat) for within-session analysis
  • labels column is the raw JSON list; reasoning, main_category, sub_category are the flattened scalar fields derived from it
  • Pre-cleaned: "8. Others" / "8.1 Others" rows are already excluded; no missing values or invalid categories

Taxonomy

See intent_classification/categories.py for the full list of valid (main_category, sub_category) pairs.