| --- |
| license: other |
| base_model: intfloat/multilingual-e5-base |
| pipeline_tag: text-classification |
| tags: [multi-label, news, multilingual] |
| --- |
| |
| # News topic classifier (e5-base) |
|
|
| Fine-tuned `intfloat/multilingual-e5-base` for multi-label classification of news articles into 13 topics: |
| Business And Finance, Market And Exchange, Demographics, Scientific Exploration, Labour Dispute, Disaster, |
| Conflicts And War, Politics, Crime, Theft, Assault, Corporate Crime, Fraud. No label means "Other". |
|
|
| ## Usage |
|
|
| - Input: `"passage: " + title + "\n" + content`, truncated to 512 tokens. |
| - Output: sigmoid per category, compared with the per-category threshold in `thresholds.json` (not 0.5). |
| - If Theft or Assault is predicted, add Crime as well. |
|
|
| ```python |
| import json |
| |
| import torch |
| from huggingface_hub import hf_hub_download |
| from transformers import AutoModelForSequenceClassification, AutoTokenizer |
| |
| REPO = "maxafanasiev/news-topic-classifier-e5-base" # private: run `hf auth login` first |
| PREFIX = "passage: " |
| PARENTS = {"Theft": "Crime", "Assault": "Crime"} |
| |
| tokenizer = AutoTokenizer.from_pretrained(REPO) |
| model = AutoModelForSequenceClassification.from_pretrained(REPO).eval() |
| thresholds = json.load(open(hf_hub_download(REPO, "thresholds.json"))) |
| categories = list(thresholds) # same order as the model outputs |
| |
| |
| @torch.no_grad() |
| def classify(articles: list[dict], max_len: int = 512) -> list[list[str]]: |
| texts = [f"{PREFIX}{a['title']}\n{a['content']}" for a in articles] |
| batch = tokenizer(texts, truncation=True, max_length=max_len, padding=True, return_tensors="pt") |
| probs = torch.sigmoid(model(**batch).logits) |
| results = [] |
| for row in probs: |
| labels = [c for c, p in zip(categories, row.tolist()) if p >= thresholds[c]] |
| labels += [PARENTS[c] for c in labels if c in PARENTS and PARENTS[c] not in labels] |
| results.append(labels or ["Other"]) |
| return results |
| |
| |
| print(classify([ |
| {"title": "Police arrest two men over jewellery store robbery", "content": "Officers detained the suspects after ..."}, |
| {"title": "El banco central sube los tipos de interés", "content": "La decisión responde a la inflación ..."}, |
| ])) |
| # e.g. [['Crime', 'Theft'], ['Business And Finance']] |
| ``` |
|
|
| Classify in batches of 16–32 articles; `max_len=256` is about twice as fast at a small quality cost (see Results). |
|
|
| ## Download once, use without Hugging Face |
|
|
| 1. Download the files (needs access to this private repo once), with any of: |
| - CLI: `hf auth login`, then `hf download maxafanasiev/news-topic-classifier-e5-base --local-dir ./news-topic-classifier-e5-base` |
| - git: `git lfs install && git clone https://huggingface.co/maxafanasiev/news-topic-classifier-e5-base` (use a token as the password) |
| - browser: the "Files and versions" tab, download every file into one folder |
| - one archive: download `news-topic-classifier-e5-base.tar.gz` from "Files and versions" (or `hf download maxafanasiev/news-topic-classifier-e5-base news-topic-classifier-e5-base.tar.gz --local-dir .`), then `tar xzf news-topic-classifier-e5-base.tar.gz`; it unpacks into `./news-topic-classifier-e5-base/`, ready for the code below. Windows 10+ has `tar` built in. |
| 2. Keep these files together: `model.safetensors`, `config.json`, `tokenizer.json`, `tokenizer_config.json`, `thresholds.json`. |
| 3. Load from the folder; `HF_HUB_OFFLINE=1` makes sure nothing is fetched from the internet: |
|
|
| ```python |
| import json |
| import os |
| from pathlib import Path |
| |
| os.environ["HF_HUB_OFFLINE"] = "1" |
| |
| import torch |
| from transformers import AutoModelForSequenceClassification, AutoTokenizer |
| |
| MODEL_DIR = Path("./news-topic-classifier-e5-base") |
| PREFIX = "passage: " |
| PARENTS = {"Theft": "Crime", "Assault": "Crime"} |
| |
| tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR) |
| model = AutoModelForSequenceClassification.from_pretrained(MODEL_DIR).eval() |
| thresholds = json.loads((MODEL_DIR / "thresholds.json").read_text()) |
| categories = list(thresholds) |
| |
| |
| @torch.no_grad() |
| def classify(articles: list[dict], max_len: int = 512) -> list[list[str]]: |
| texts = [f"{PREFIX}{a['title']}\n{a['content']}" for a in articles] |
| batch = tokenizer(texts, truncation=True, max_length=max_len, padding=True, return_tensors="pt") |
| probs = torch.sigmoid(model(**batch).logits) |
| results = [] |
| for row in probs: |
| labels = [c for c, p in zip(categories, row.tolist()) if p >= thresholds[c]] |
| labels += [PARENTS[c] for c in labels if c in PARENTS and PARENTS[c] not in labels] |
| results.append(labels or ["Other"]) |
| return results |
| ``` |
|
|
| Requirements: `torch` and `transformers` (no `huggingface_hub` login needed at this point). |
|
|
| ## Results (gold set, CPU Apple M4 Pro, fp32) |
|
|
| | Max tokens | micro-F1 (random gold, 163) | exact set match | CPU latency | |
| | --- | --- | --- | --- | |
| | 512 | 0.860 | 0.804 | 42 ms/article | |
| | 256 | 0.836 | 0.785 | 20 ms/article | |
|
|
| Training data: ~8k news articles labeled by an LLM teacher (Sonnet 5.5); gold built from three blind LLM labelers plus a judge, not hand-checked. Training data is private and not included. |
|
|
