Meeting transcript coding pipeline

A research pipeline for assigning emotional-intelligence and team-interaction codes to meeting transcript turns. Each nonempty turn is evaluated independently for each of 21 labels, using codebook definitions, examples, preceding turns and optional retrieved examples.

This is a sanitized collaborator repository, not the original research archive. It has fresh Git history. It includes implementation, synthetic tests, aggregate development results and documentation. No participant transcript workbook, retrieval index, prompt ledger or prediction workbook is distributed.

The bundled codebook examples are fictional replacements. They do not reproduce the reported study results. Authorized study reproduction requires the frozen private workbook and codebook, plus the recorded model/runtime and original source version. See reproduction.

Start here

• Project overview and implementation
• Data split and evaluation methods
• Tests run and actual results
• Installation and reproduction
• Data privacy and sharing boundaries

Quick start: no private data, no GPU, no network inference

Install Python 3.12+ and uv, then run from the repository root:

uv sync --frozen
uv run --frozen pytest -q
uv run --frozen python scripts/synthetic_demo.py
uv run --frozen ptg --help
uv run --frozen python scripts/check_repository.py

The synthetic demo exercises the real workbook ingestion, FAISS retrieval, binary prompting, checkpoints, export and evaluation. It replaces only embedding/model network calls with deterministic test doubles. Its scores test plumbing, not LLM quality. Generated files remain in ignored artifacts/.

Real inference

The code needs an Ollama embedding service and an OpenAI-compatible chat service, supplied separately. uv sync does not install models or start a model server. The old variable name LM_STUDIOS_HOST is retained for compatibility; LM Studio is not mandatory.

cp .env.example .env
# Set approved local endpoints and exact model identifiers.
# Supply an authorized workbook privately, never commit it.
uv run --frozen ptg --config config/study-21.yaml build-index --semester F25 --teams 1,2 --meets 1,2
uv run --frozen ptg --config config/study-21.yaml infer --semester S26 --team 1 --meet 1 --pilot-row-limit 10 --lm-workers 2
uv run --frozen ptg --config config/study-21.yaml export --semester S26 --team 1 --meet 1 --pilot 10

These commands with the synthetic template codebook create a new exploratory run, not a historical reproduction. Follow the private-config instructions in the reproduction guide for the research codebook. Index and target sheets must differ.

Current evidence

• September 24, 2026: Qwen and Granite runs on 1,256 reconciled development turns, all 21 labels. Micro-F1: Qwen 0.2398, Granite 0.3628.
• Granite's descriptive mean of 42 sheet-label Cohen's kappas: 0.2372, against reconciled codes, excluding its one invalid decision for kappa only.
• September 29: controlled local Qwen screen on 128 development turns. Removing dynamic retrieval improved micro-F1 from 0.2011 to 0.2615.
• These are development findings. They do not demonstrate human-level coding or independent final-test performance. Human-human agreement on confirmed independently coded rows remains to be established.

Layout

src/ptg/        Ingest, index, retrieve, classify, checkpoint, export
scripts/        Audits, scoring, controlled experiments, synthetic demo
config/         Model settings and clearly marked synthetic codebooks
notebooks/      Optional opt-in Colab staging, not required for local use
tests/          Synthetic regression tests, no participant fixtures
docs/           Collaborator-facing project and evaluation documentation
results/        Aggregate-only metrics and per-label confusion counts

config/study-23.yaml and its extended codebook are archived compatibility fixtures, not an approved evaluated arm. Optional rental/watchdog scripts are retained for code review but require separately configured approval, credentials and safety timers. No cloud processing is authorized by this repository.

Repository access

FloorIsGround/ptg-classifier is a private repository. Participant transcripts, private reproduction inputs and original research history are not included.

Before wider sharing or public release, review construct-definition redistribution permissions, aggregate disclosure and collaborator access. No redistribution license is asserted until the project owner/institution confirms it. See data privacy and sharing boundaries.