Financial Topic Classifier — Qwen2.5-0.5B + LoRA

A LoRA adapter for Qwen/Qwen2.5-0.5B that classifies short English financial texts (tweets) into one of 20 topics. It trains 1,099,264 parameters (0.22% of the base model): LoRA on the attention projections plus a linear classification head on the last token.

Labels

0 Analyst Update, 1 Fed | Central Banks, 2 Company | Product News, 3 Treasuries | Corporate Debt, 4 Dividend, 5 Earnings, 6 Energy | Oil, 7 Financials, 8 Currencies, 9 General News | Opinion, 10 Gold | Metals | Materials, 11 IPO, 12 Legal | Regulation, 13 M&A | Investments, 14 Macro, 15 Markets, 16 Politics, 17 Personnel Change, 18 Stock Commentary, 19 Stock Movement

Results

Test set: the dataset's official validation split (4,117 tweets), evaluated once after training. Approximate 95% margin on accuracy: ±0.9 points.

On the 3,439 test tweets whose text never appears in the training data: accuracy 0.9023, macro F1 0.8842.

Training data

zeroshot/twitter-financial-news-topic (MIT license). The original train split was divided 90/10, stratified by label (seed 42), into train (15,291) and dev (1,699). Dev was used for checkpoint selection and to fit the calibration temperature. Largest/smallest class ratio in train: 81x.

Training procedure

• LoRA: r=8, alpha=16, dropout=0.05, target modules: k_proj, o_proj, q_proj, v_proj; classification head trained fully
• Max length 96 tokens with truncation; dynamic padding
• 3 epochs, learning rate 0.0002, AdamW, linear schedule with 71 warmup steps
• Effective batch size 32 (16 per device x 2 GPU)
• Precision: fp16 mixed precision
• Hardware: Tesla T4, Tesla T4; training time 26 min; peak GPU memory 6.0 GB
• Best checkpoint chosen by dev macro F1 (epoch 3)

Most confused topic pairs (test set)

Calibration

Temperature scaling with T = 1.6 (fitted on dev) changes test expected calibration error from 0.043 to 0.011 without changing any prediction. Divide the logits by T before softmax when using confidence scores.

Limitations

• Trained on English financial tweets only; other domains, languages and long documents are untested.
• Several topics overlap in meaning (see the confused pairs), and each tweet carries exactly one label even when it covers two topics.
• Rare classes have few test examples (e.g. Gold | Metals | Materials: 13), so their per-class scores are noisy.
• Labels are used as provided by the dataset; some appear noisy.

How to use

import json, torch
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel

repo = "ZainabAlmahal/financial-topic-classifier-lora"
labels = json.load(open(hf_hub_download(repo, "label_names.json")))
tokenizer = AutoTokenizer.from_pretrained(repo)
base = AutoModelForSequenceClassification.from_pretrained("Qwen/Qwen2.5-0.5B", num_labels=20)
base.config.pad_token_id = tokenizer.pad_token_id
model = PeftModel.from_pretrained(base, repo).merge_and_unload().eval()

inputs = tokenizer(["Apple beats quarterly earnings estimates"], truncation=True, max_length=96, return_tensors="pt")
with torch.no_grad():
    probs = torch.softmax(model(**inputs).logits / 1.6, dim=-1)   # 1.6 = calibration temperature
print(labels[probs.argmax(-1).item()], round(probs.max().item(), 3))