scripts/fix_all_o_ratio.py

#!/usr/bin/env python3
"""Downsample all-O (no entity) sentences to ~10% of training/valid data."""

import json
import random
from collections import Counter
from pathlib import Path

SEED = 42
TARGET_RATIO = 0.10  # 10% all-O


def has_entities(example: dict) -> bool:
    spans = example.get("spans", {})
    return any(len(v) > 0 for v in spans.values())


def process_file(input_path: Path, output_path: Path):
    print(f"\n{'='*60}")
    print(f"Processing: {input_path.name}")
    print(f"{'='*60}")

    with open(input_path) as f:
        data = [json.loads(line) for line in f]

    entity_examples = [ex for ex in data if has_entities(ex)]
    all_o_examples = [ex for ex in data if not has_entities(ex)]

    n_entity = len(entity_examples)
    n_all_o = len(all_o_examples)
    orig_ratio = n_all_o / len(data)

    print(f"Original: {len(data)} total, {n_entity} entity, {n_all_o} all-O ({orig_ratio:.1%})")

    # Source distribution of all-O
    source_counts = Counter(ex.get("info", {}).get("source", "unknown") for ex in all_o_examples)
    print(f"\nAll-O by source:")
    for src, cnt in source_counts.most_common():
        print(f"  {src}: {cnt}")

    # Target: N_entity / 9 all-O examples → 10% of total
    target_all_o = round(n_entity / 9)
    print(f"\nTarget all-O count: {target_all_o}")

    if target_all_o >= n_all_o:
        print("Already at or below target — no downsampling needed.")
        kept = all_o_examples
    else:
        # Proportional downsampling by source
        rng = random.Random(SEED)
        kept = []
        for src, cnt in source_counts.items():
            proportion = cnt / n_all_o
            n_keep = max(1, round(target_all_o * proportion))
            src_examples = [ex for ex in all_o_examples if ex.get("info", {}).get("source", "unknown") == src]
            rng.shuffle(src_examples)
            kept.extend(src_examples[:n_keep])

        # Trim or pad to exact target
        if len(kept) > target_all_o:
            rng.shuffle(kept)
            kept = kept[:target_all_o]

    # Combine and preserve original order by re-scanning
    kept_ids = {ex.get("info", {}).get("id") for ex in kept}
    result = [ex for ex in data if has_entities(ex) or ex.get("info", {}).get("id") in kept_ids]

    final_all_o = sum(1 for ex in result if not has_entities(ex))
    final_ratio = final_all_o / len(result)

    print(f"\nResult: {len(result)} total, {len(result) - final_all_o} entity, {final_all_o} all-O ({final_ratio:.1%})")
    print(f"Removed: {n_all_o - final_all_o} all-O examples")

    with open(output_path, "w") as f:
        for ex in result:
            f.write(json.dumps(ex, ensure_ascii=False) + "\n")

    print(f"Written: {output_path}")


if __name__ == "__main__":
    base = Path("/home/ubuntu/alkyline/data/processed")

    process_file(
        base / "enriched_5class_train_cleaned.jsonl",
        base / "enriched_5class_train_cleaned_trimmed.jsonl",
    )
    process_file(
        base / "enriched_5class_valid_cleaned.jsonl",
        base / "enriched_5class_valid_cleaned_trimmed.jsonl",
    )