stage4-generation.md

Stage 4: Turn-taking dialogue generation

Queries the predictor at each slot and inserts the chosen behavior, then filters role-confused outputs.

This stage regenerates each dialogue turn by turn and, at every slot Stage 2 detects inside a user turn, samples a turn-taking behavior from the predictor's distribution. It drives three separate LLM roles, each with its own model flag, so decide what serves each one before running:

Role	Flag	Default	What it does
Writer	--llm_model_name (--api_key, --base_url)	gpt-4.1	Generates each turn's text and judges when the dialogue is done
Boundary detector (Stage 2)	--boundary_model_name (--boundary_api_key, --boundary_base_url)	gpt-4.1-mini	Marks clause boundaries, which become the candidate slots
Turn-taking predictor (Stage 3)	--tt_model_name (--tt_api_key, --tt_base_url)	Qwen/Qwen3-14B	Scores floor_taking / backchannel / silence at each slot

Each follows the model-name rule from Setup independently, so the defaults mean a bare run needs both OPENAI_API_KEY and a self-hosted Qwen/Qwen3-14B at http://localhost:8000/v1. Name every endpoint explicitly:

export OPENAI_API_KEY=sk-...
.venv/bin/python -m src.synthesis.run \
  --dataset interviewer --split train \
  --input_root data-dialogues/ \
  --save_root outputs/generated_dialogues_with_tt \
  --llm_model_name gpt-4.1 --boundary_model_name gpt-4.1-mini \
  --tt_model_name Qwen/Qwen3-32B --tt_base_url http://localhost:8000/v1

Using the Stage 3 predictor instead

--tt_model_name asks a chat model for verbalized probabilities. To use the LoRA token classifier Stage 3 trains — the configuration behind the released corpus, whose dialogues carry "tt_mode": "hf_classification" — pass --hf_model_name_or_path instead. It takes priority over --tt_model_name, needs no --tt_base_url, and loads in process, so this run needs a GPU:

.venv/bin/python -m src.synthesis.run \
  --dataset interviewer --split train \
  --input_root data-dialogues/ \
  --save_root outputs/generated_dialogues_with_tt \
  --llm_model_name gpt-4.1 --boundary_model_name gpt-4.1-mini \
  --hf_model_name_or_path ./output/turntaking_qwen3_4b --hf_load_in_4bit

Point it at the --output_dir of src.train_turntaking_hf; the base model is read from the adapter's adapter_config.json, so there is no separate --model_name_or_path here.

Inputs and outputs

--dataset accepts socraticlm | multiwoz | interviewer | negotiator | persuader | soda | all; --split is train | test. --input_root is required and expects the text_dialogue_<dataset>/<split>/*.json layout produced by src/prepare_corpus.py — not the published HF layout; convert with tools/unpack_corpus.py --kind dialogues first (see CORPUS.md). That half is train-only, so --split test needs dialogues of your own from Stage 1. Results are written to <save_root>/text_dialogue_<dataset>/<split>/*.json, and an existing output file is skipped, so a re-run resumes. --max_dialogues (1000) and --max_turns (20) bound the cost of a trial run.

Backchannel decisions are recorded per slot in each turn's history, not inserted into content: only [TAKE_FLOOR] appears in the transcript text. This matches the released corpus, and Stage 5 renders backchannels from the slot metadata.

Backchannel content

Optional, and comes last, filling in what each backchannel slot actually says. It is a fourth endpoint, defaulting to Qwen/Qwen3-14B at http://localhost:8008/v1 — port 8008, not the 8000 the other stages default to, so that it can be a second server:

.venv/bin/python -m src.synthesis.run_add_bc \
  --input_root outputs/generated_dialogues_with_tt \
  --output_root outputs/generated_dialogues_with_hf_swbd_plus_backchannels \
  --model_name Qwen/Qwen3-32B --base_url http://localhost:8000/v1

Both roots are required. It writes the generated text into each backchannel slot's content field, leaving the transcript text unchanged.

Source: src/synthesis/*.py.

Next: Stage 5.