script/self_redundancy.py

import json
import re
import numpy as np
from sentence_transformers import SentenceTransformer
from nltk.tokenize import sent_tokenize
from pathlib import Path

input_path = Path("/root/test/weitiao/data_process_bq/data/train2_filtered_by_length_replaced.json")
output_path = Path("/root/test/weitiao/data_process_bq/data/train2_stage1_filtered.json")

model = SentenceTransformer("/root/test/weitiao/data_process_bq/model/all-MiniLM-L6-v2")

def normalize(text):
    text = text.lower()
    text = re.sub(r"[^\w\s]", "", text)
    return text

def jaccard_overlap(a, b):
    a_set = set(normalize(a).split())
    b_set = set(normalize(b).split())
    if len(a_set) == 0:
        return 0.0
    return len(a_set & b_set) / len(a_set)

def sentence_redundancy(text):
    sents = sent_tokenize(text)
    if len(sents) <= 1:
        return 0.0

    emb = model.encode(sents, normalize_embeddings=True)
    sims = emb @ emb.T
    upper = sims[np.triu_indices(len(sents), k=1)]
    return float(upper.max()) if len(upper) > 0 else 0.0

def semantic_sim(a, b):
    emb = model.encode([a, b], normalize_embeddings=True)
    return float(emb[0] @ emb[1])

def stage1_filter_one(sample,
                      overlap_th=0.35,
                      redundancy_th=0.9,
                      delta_s_th=0.88):

    messages = sample["messages"]
    chosen_resp = sample["chosen"][0]["content"]

    prompt_parts = []
    prev_response = None

    for m in messages:
        if m["role"] == "assistant":
            prev_response = m["content"]
        prompt_parts.append(f'{m["role"]}: {m["content"]}')

    prompt = "\n".join(prompt_parts)

    # ① 高词面重叠
    overlap = jaccard_overlap(chosen_resp, prompt)
    if overlap > overlap_th:
        return False, "HIGH_PROMPT_OVERLAP"

    # ② 自身重复
    redundancy = sentence_redundancy(chosen_resp)
    if redundancy > redundancy_th:
        return False, "HIGH_SELF_REDUNDANCY"

    # ③ ΔS ≈ 0（与上一轮 assistant 几乎一样）
    if prev_response is not None:
        ds = semantic_sim(prev_response, chosen_resp)
        if ds > delta_s_th:
            return False, "DELTA_S_ZERO"

    return True, "KEEP"

with open(input_path, "r", encoding="utf-8") as f:
    data = json.load(f)

kept = []
stats = {}

for sample in data:
    keep, reason = stage1_filter_one(sample)
    stats[reason] = stats.get(reason, 0) + 1
    if keep:
        kept.append(sample)

with open(output_path, "w", encoding="utf-8") as f:
    json.dump(kept, f, ensure_ascii=False, indent=2)

print("Stage 1 stats:")
for k, v in stats.items():
    print(f"{k}: {v}")
print(f"Kept {len(kept)} / {len(data)}")