Thaki-BanCode-KoEn-v2 / bancode.py
thaki-AI's picture
Release Thaki-BanCode-KoEn-v2 (weights, config, wrapper; tokenizer loaded from upstream)
a818f1f verified
Raw History Blame Contribute Delete
2.54 kB
# SPDX-License-Identifier: MIT
# Copyright (c) 2026 ThakiCloud
"""BanCode v2 inference wrapper (CPU or GPU).
The tokenizer is NOT part of this repository. It is loaded from jhu-clsp/mmBERT-small at the exact
revision pinned in bancode_v2_profile.json and keeps its own terms (see NOTICE.md).
python bancode.py <model_dir_or_repo_id> "text one" "text two"
"""
import json
import os
import sys
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
PROFILE = "bancode_v2_profile.json"
def _profile_path(model, revision=None):
if os.path.isdir(model):
return os.path.join(model, PROFILE)
from huggingface_hub import hf_hub_download
return hf_hub_download(model, PROFILE, revision=revision)
class BanCodeV2:
def __init__(self, model=".", revision=None, device="cpu"):
with open(_profile_path(model, revision), encoding="utf-8") as f:
self.profile = json.load(f)
t = self.profile["tokenizer"]
self.tokenizer = AutoTokenizer.from_pretrained(t["repo"], revision=t["revision"])
self.model = AutoModelForSequenceClassification.from_pretrained(model, revision=revision).eval().to(device)
self.device = device
self.code_idx = int(self.model.config.label2id[self.profile["code_label"]])
def logits(self, texts):
enc = self.tokenizer(list(texts), truncation=True, max_length=self.profile["max_length"],
padding=True, return_tensors="pt").to(self.device)
with torch.inference_mode():
return self.model(**enc).logits.float()
def p_code(self, texts):
return torch.softmax(self.logits(texts), -1)[:, self.code_idx].tolist()
def scan(self, prompt):
"""LLM Guard scanner contract: (sanitized_prompt, is_valid, risk_score). Raw text, no preprocessing."""
if prompt.strip() == "":
return prompt, True, -1.0
s, t = self.p_code([prompt])[0], self.profile["threshold"]
risk = round((s - t) / (1 - t), 1) if s > t else round((s - t) / t, 1)
return prompt, not (s > t), min(max(risk, -1.0), 1.0)
if __name__ == "__main__":
m = BanCodeV2(sys.argv[1] if len(sys.argv) > 1 else ".")
for text in sys.argv[2:] or ["Please summarise the meeting notes from Thursday.",
"Run this:\n```bash\npip install requests\n```"]:
_, ok, risk = m.scan(text)
print(json.dumps({"code_detected": not ok, "risk_score": risk, "text": text[:40]}, ensure_ascii=False))