A LoRA adapter for google/gemma-3-270m-it, fine-tuned as an ultra-lightweight text-only baseline for detecting financial scams in short-form social video (YouTube + TikTok), with a Philippine / Filipino focus. Given a video's transcript + title + description, it predicts scam vs. legitimate using the OptiScam C1-C7 policy criteria.

At 270M parameters this is the smallest member of the Filipino Scam Detection model family - a size / latency baseline. For real use, prefer the multimodal Scam-Qwen3-VL-RsLoRA (F1 0.804).

Model Details

Developed by: Jules Gregory R. Agustin (TamAko783)
Model type: Text classifier / reasoner - LoRA adapter (PEFT)
Task: Scam vs. not-scam detection from a video's text channels
Language(s): English, Filipino / Tagalog (code-switched)
Finetuned from: google/gemma-3-270m-it
Base model license: Gemma Terms of Use. Adapter license: see repository.

Training Data

Trained on the OptiScam: Multimodal Filipino Scam Video Dataset (https://www.kaggle.com/datasets/julesgregoryagustin/training-filipino-scam-dataset), ~2,000 YouTube/TikTok videos balanced 1,000 scam / 1,000 legitimate. Videos were scraped with Playwright and Selenium; speech was transcribed with Whisper; labels and rationales were produced by teacher LLM labelers (Google Gemma and Google Gemini) against the C1-C7 criteria. Only the transcript + title + description channels are used.

Detection criteria (C1-C7)

Labels follow the OptiScam C1-C7 policy scheme, adapted from the YouTube-policy criteria in Kulsum et al. (see Citation). The criteria a video matches are stored alongside each example.

C1 Criminal claim - promotes illegal activity or stolen credentials as a money path.
C2 Unbounded giveaway - free cash, phones, or load with no sponsor, qualifier, or limit.
C3 Off-site redirect - pushes viewers to Telegram, WhatsApp, sketchy domains, or DM flows.
C4 Fast-money lure - promises large returns with little effort or skill in a short window.
C5 Harmful link - phishing URL, fake login page, or OTP / PIN / seed-phrase requests.
C6 Get-rich-quick - mobile mining, task apps with guaranteed earnings, or no-loss trading bots.
C7 Impersonation - mimics a real bank, celebrity, government agency, or known brand.

Extended criteria (Filipino scam archetypes): P2E (Play-to-Earn), Task (Telegram / Group Task), and E-Wallet (GCash / Maya phishing), alongside the global types (crypto, gift card, giveaway).

Evaluation

Held-out test split:

Metric	Value
Accuracy	0.505
F1	0.576

At 270M parameters this baseline performs near chance on accuracy - it is published for size/latency comparison within the model family, not for production use. The multimodal Qwen3-VL model (F1 0.804) and the Qwen2.5-1.5B text model (F1 0.703) are substantially stronger.

Uses

Direct use: research / educational baseline; low-resource on-device experimentation.
Out-of-scope: production scam enforcement, non-financial-scam domains, or any use treating its output as authoritative. Keep a human in the loop.

How to Get Started

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "google/gemma-3-270m-it"
ADAPTER = "TamAko783/Scam_Gemma3-270M-IT_Text_epochFinal"

tok = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, ADAPTER).eval()

SYSTEM = "You are a scam-detection analyst. Decide if the video is a scam and cite criteria (C1-C7)."
user = "Transcript: ...\nTitle: ...\nDescription: ...\n\nIs this a scam video? Explain with C1-C7."
msgs = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": user}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0, inputs.shape[1]:], skip_special_tokens=True))

Citation

Training data and inspiration - the C1-C7 policy criteria are adapted from:

Ummay Kulsum, Aafaq Sabir, Abhinaya S.B., and Anupam Das. "Beyond Metadata: Multimodal, Policy-Aware Detection of YouTube Scam Videos." ICWSM 2026. arXiv:2509.23418.

bibtex

@inproceedings{kulsum2026beyond,
  title     = {Beyond Metadata: Multimodal, Policy-Aware Detection of YouTube Scam Videos},
  author    = {Kulsum, Ummay and Sabir, Aafaq and Abhinaya, S.B. and Das, Anupam},
  booktitle = {Proceedings of the International AAAI Conference on Web and Social Media (ICWSM)},
  year      = {2026},
  note      = {arXiv:2509.23418}
}

Dataset: OptiScam: Multimodal Filipino Scam Video Dataset - https://www.kaggle.com/datasets/julesgregoryagustin/training-filipino-scam-dataset

Model Card Authors

Jules Gregory R. Agustin (TamAko783)