ademchaoua/freelance-offer-request-classifier

Freelance offer/request classifier

Fine-tuned all-MiniLM-L6-v2 (22M params) that classifies short freelance-community messages into one of three categories:

• offer — the sender is offering their own service/skill
• request — the sender is looking to hire / needs someone else's service
• neither — general chat, unrelated to offering/requesting a service

Performance

Evaluated on a held-out validation split (15% of training data, not seen during training):

Class	Precision	Recall	F1
offer	0.74	0.87	0.80
request	0.93	0.88	0.90
neither	0.86	0.78	0.82
accuracy			0.85
macro avg	0.84	0.84	0.84

Trained on 2,614 messages collected from freelance-community Telegram groups.

Known limitation

The model still struggles with the phrasing pattern "I'm looking for [role] opportunities" when the speaker is actually offering their own skill (it tends to predict "request" instead of "offer" for this pattern with high confidence). If your use case is sensitive to this, consider a post-processing rule for this specific phrasing, or contribute additional labeled examples.

Usage (ONNX, recommended — fast, CPU-only)

from optimum.onnxruntime import ORTModelForSequenceClassification
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("ademchaoua/freelance-offer-request-classifier", subfolder="onnx")
model = ORTModelForSequenceClassification.from_pretrained("ademchaoua/freelance-offer-request-classifier", subfolder="onnx")

inputs = tokenizer("I offer web scraping services", return_tensors="pt")
outputs = model(**inputs)
pred = outputs.logits.argmax(-1).item()
print(model.config.id2label[pred])

Usage (PyTorch)

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("ademchaoua/freelance-offer-request-classifier")
model = AutoModelForSequenceClassification.from_pretrained("ademchaoua/freelance-offer-request-classifier")

inputs = tokenizer("Need a copywriter for email sequences", return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
pred = logits.argmax(-1).item()
print(model.config.id2label[pred])

Training details

• Base model: sentence-transformers/all-MiniLM-L6-v2
• Method: end-to-end fine-tuning (full model, classification head included)
• Max sequence length: 96 tokens
• Class-weighted loss (balanced) to counter class imbalance
• Early stopping on macro-F1
• Trained on ~2,600 messages collected from freelance-community Telegram groups, labeled with an LLM (DeepSeek) and manually reviewed
• Exported to ONNX + INT8 quantized for fast CPU inference (~2ms/text)