scripts/build_conversations.py

#!/usr/bin/env python3
"""Build SFT conversation pairs from parsed transcript turns.

Reads dataset/data/turns.jsonl and produces:
  - dataset/data/conversations.jsonl  (ShareGPT multi-turn format)
  - dataset/data/sft_pairs.jsonl      (individual instruction/response pairs)
"""

import json
import sys
from collections import defaultdict
from pathlib import Path

DATA_DIR = Path(__file__).resolve().parent.parent / "data"
TURNS_FILE = DATA_DIR / "turns.jsonl"
CONV_FILE = DATA_DIR / "conversations.jsonl"
SFT_FILE = DATA_DIR / "sft_pairs.jsonl"

MIN_RESPONSE_LEN = 20  # Skip trivially short Audrey responses in SFT pairs
MAX_CONTEXT_TURNS = 3  # Max preceding turns to include as instruction context


def load_turns_by_file(path: Path) -> dict:
    """Load turns grouped by source_file, preserving turn_index order."""
    groups = defaultdict(list)
    with open(path, encoding="utf-8") as f:
        for line in f:
            turn = json.loads(line)
            groups[turn["source_file"]].append(turn)
    # Sort each group by turn_index
    for sf in groups:
        groups[sf].sort(key=lambda t: t["turn_index"])
    return groups


def merge_messages(turns: list[dict]) -> list[dict]:
    """Merge consecutive same-role turns into ShareGPT messages.

    Audrey → role "assistant", everyone else → role "human".
    Consecutive human turns from different speakers get speaker attribution.
    Consecutive assistant turns are concatenated.
    """
    messages = []
    for turn in turns:
        role = "assistant" if turn["is_audrey"] else "human"
        speaker = turn["speaker"]
        text = turn["text"].strip()
        if not text:
            continue

        if messages and messages[-1]["role"] == role:
            # Merge into previous message
            prev = messages[-1]
            if role == "human":
                # Add speaker attribution to both previous and current if needed
                if prev.get("_multi_speaker") is False:
                    # First merge: retrofit attribution on the previous text
                    prev["content"] = f"**{prev['speaker']}:** {prev['content']}"
                    prev["_multi_speaker"] = True
                prev["content"] += f"\n\n**{speaker}:** {text}"
                prev["_multi_speaker"] = True
                # Track all speakers
                if speaker not in prev["_speakers"]:
                    prev["_speakers"].append(speaker)
            else:
                # Consecutive Audrey turns: just concatenate
                prev["content"] += f"\n\n{text}"
        else:
            messages.append({
                "role": role,
                "speaker": speaker,
                "content": text,
                "_multi_speaker": False,
                "_speakers": [speaker],
            })

    # Clean up internal tracking fields
    for msg in messages:
        if msg["_multi_speaker"]:
            # Use all speakers joined
            msg["speaker"] = ", ".join(msg["_speakers"])
        del msg["_multi_speaker"]
        del msg["_speakers"]

    return messages


def build_conversation(source_file: str, turns: list[dict]) -> dict:
    """Build a ShareGPT-format conversation record."""
    first = turns[0]
    conv_id = source_file.replace(".md", "")
    messages = merge_messages(turns)

    # Collect all unique non-Audrey participants
    participants = []
    seen = set()
    for t in turns:
        if t["speaker"] not in seen:
            seen.add(t["speaker"])
            participants.append(t["speaker"])

    audrey_word_count = sum(
        len(m["content"].split()) if first["language"] == "en"
        else len(m["content"])
        for m in messages if m["role"] == "assistant"
    )

    return {
        "id": conv_id,
        "date": first["date"],
        "title": first["title"],
        "language": first["language"],
        "participants": participants,
        "messages": [{"role": m["role"], "speaker": m["speaker"], "content": m["content"]} for m in messages],
        "source_file": source_file,
        "num_turns": len(messages),
        "audrey_word_count": audrey_word_count,
    }


def build_sft_pairs_from_conversation(conv: dict) -> list[dict]:
    """Extract SFT pairs from a multi-turn conversation.

    For each assistant turn, include up to MAX_CONTEXT_TURNS preceding turns
    as instruction context.
    """
    pairs = []
    messages = conv["messages"]
    pair_idx = 0

    for i, msg in enumerate(messages):
        if msg["role"] != "assistant":
            continue
        response = msg["content"]
        if len(response) < MIN_RESPONSE_LEN:
            continue

        # Gather preceding context (up to MAX_CONTEXT_TURNS)
        context_start = max(0, i - MAX_CONTEXT_TURNS)
        context_msgs = messages[context_start:i]

        if not context_msgs:
            # Assistant speaks first with no preceding context — skip for SFT
            continue

        # Build instruction from context turns
        instruction_parts = []
        for cm in context_msgs:
            if cm["role"] == "human":
                instruction_parts.append(cm["content"])
            else:
                # Prior assistant turn as context
                instruction_parts.append(f"[Audrey Tang:] {cm['content']}")
        instruction = "\n\n".join(instruction_parts)

        pairs.append({
            "id": f"{conv['id']}/pair_{pair_idx:03d}",
            "date": conv["date"],
            "title": conv["title"],
            "language": conv["language"],
            "instruction": instruction,
            "response": response,
            "context_turns": len(context_msgs),
            "source_file": conv["source_file"],
        })
        pair_idx += 1

    return pairs


def build_sft_pairs_from_monologue(conv: dict) -> list[dict]:
    """Build SFT pairs from a monologue (Audrey-only single-speaker file).

    Split the text into paragraphs and use the title as instruction.
    """
    pairs = []
    messages = conv["messages"]
    if not messages:
        return pairs

    # Combine all message content (should be one merged assistant message)
    full_text = "\n\n".join(m["content"] for m in messages)

    # Split into paragraph groups (~500 chars each for reasonable SFT chunks)
    paragraphs = [p.strip() for p in full_text.split("\n\n") if p.strip()]
    if not paragraphs:
        return pairs

    # If the whole text is short enough, make one pair
    if len(paragraphs) <= 3 or len(full_text) < 500:
        if len(full_text) >= MIN_RESPONSE_LEN:
            pairs.append({
                "id": f"{conv['id']}/monologue_000",
                "date": conv["date"],
                "title": conv["title"],
                "language": conv["language"],
                "instruction": conv["title"],
                "response": full_text,
                "context_turns": 0,
                "source_file": conv["source_file"],
            })
        return pairs

    # Group paragraphs into chunks
    chunks = []
    current_chunk = []
    current_len = 0
    for para in paragraphs:
        current_chunk.append(para)
        current_len += len(para)
        if current_len >= 400:
            chunks.append("\n\n".join(current_chunk))
            current_chunk = []
            current_len = 0
    if current_chunk:
        chunks.append("\n\n".join(current_chunk))

    for idx, chunk in enumerate(chunks):
        if len(chunk) < MIN_RESPONSE_LEN:
            continue
        if idx == 0:
            instruction = conv["title"]
        else:
            instruction = f"{conv['title']} (continued)"
        pairs.append({
            "id": f"{conv['id']}/monologue_{idx:03d}",
            "date": conv["date"],
            "title": conv["title"],
            "language": conv["language"],
            "instruction": instruction,
            "response": chunk,
            "context_turns": 0,
            "source_file": conv["source_file"],
        })

    return pairs


def is_monologue(turns: list[dict]) -> bool:
    """Check if all turns are from a single speaker who is Audrey."""
    speakers = {t["speaker"] for t in turns}
    return len(speakers) == 1 and turns[0]["is_audrey"]


def main():
    print(f"Reading turns from {TURNS_FILE}...")
    groups = load_turns_by_file(TURNS_FILE)
    print(f"Found {len(groups)} source files, {sum(len(v) for v in groups.values())} total turns")

    conversations = []
    all_sft_pairs = []
    monologue_count = 0
    no_audrey_count = 0
    lang_counts = defaultdict(int)
    lang_sft_counts = defaultdict(int)

    for source_file, turns in sorted(groups.items()):
        conv = build_conversation(source_file, turns)
        conversations.append(conv)
        lang_counts[conv["language"]] += 1

        has_audrey = any(t["is_audrey"] for t in turns)

        if is_monologue(turns):
            monologue_count += 1
            pairs = build_sft_pairs_from_monologue(conv)
        elif has_audrey:
            pairs = build_sft_pairs_from_conversation(conv)
        else:
            # No Audrey in this conversation — no SFT pairs
            no_audrey_count += 1
            pairs = []

        for p in pairs:
            lang_sft_counts[p["language"]] += 1
        all_sft_pairs.extend(pairs)

    # Write outputs
    print(f"\nWriting {len(conversations)} conversations to {CONV_FILE}...")
    with open(CONV_FILE, "w", encoding="utf-8") as f:
        for conv in conversations:
            f.write(json.dumps(conv, ensure_ascii=False) + "\n")

    print(f"Writing {len(all_sft_pairs)} SFT pairs to {SFT_FILE}...")
    with open(SFT_FILE, "w", encoding="utf-8") as f:
        for pair in all_sft_pairs:
            f.write(json.dumps(pair, ensure_ascii=False) + "\n")

    # Statistics
    total_turns = sum(c["num_turns"] for c in conversations)
    avg_turns = total_turns / len(conversations) if conversations else 0
    total_audrey_wc = sum(c["audrey_word_count"] for c in conversations)

    print("\n" + "=" * 60)
    print("OUTPUT STATISTICS")
    print("=" * 60)
    print(f"Total conversations:       {len(conversations)}")
    print(f"  Monologues (Audrey only): {monologue_count}")
    print(f"  No Audrey present:        {no_audrey_count}")
    print(f"  Multi-speaker w/ Audrey:  {len(conversations) - monologue_count - no_audrey_count}")
    print(f"Total SFT pairs:           {len(all_sft_pairs)}")
    print(f"Avg turns per conversation: {avg_turns:.1f}")
    print(f"Total Audrey word/char ct:  {total_audrey_wc}")
    print()
    print("Language split (conversations):")
    for lang, count in sorted(lang_counts.items()):
        print(f"  {lang}: {count}")
    print("Language split (SFT pairs):")
    for lang, count in sorted(lang_sft_counts.items()):
        print(f"  {lang}: {count}")
    print("=" * 60)


if __name__ == "__main__":
    main()