Datasetpublic

TamAko783/filipino-scam-final-shards

Filipino Scam — Final Shards CoT-formatted pickle shards for Qwen3-VL-4B-Instruct LoRA fine-tuning on Filipino short-form video scam detection. This is the final, training-ready format used to build TamAko783/Scam-Qwen3-VL-4B-final-lora. Splits File Samples Purpose Training.pkl 1600 Training set Validate.pkl 198 Validation (early stopping + best-checkpoint selection) Evaluate.pkl 202 Held-out test (final unbiased metrics) Both Validate and… See the full description on the dataset page: https://huggingface.co/datasets/TamAko783/filipino-scam-final-shards.

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes338downloads

Dataset Card

Filipino Scam — Final Shards

CoT-formatted pickle shards for Qwen3-VL-4B-Instruct LoRA fine-tuning on Filipino short-form video scam detection. This is the final, training-ready format used to build TamAko783/Scam-Qwen3-VL-4B-final-lora.

Splits

File	Samples	Purpose
`Training.pkl`	1600	Training set
`Validate.pkl`	198	Validation (early stopping + best-checkpoint selection)
`Evaluate.pkl`	202	Held-out test (final unbiased metrics)

Both Validate and Evaluate are stratified halves of the original Kaggle test split, balanced 50/50 yes/no within every category (crypto, ewallet, giftcard, giveaway, p2e, task) using seed=42.

Format (one `pickle.dump` per sample)

python

{
  "id":       "youtube_<id>",        # original video id
  "platform": "tiktok|youtube|fb",
  "category": "crypto|ewallet|giftcard|giveaway|p2e|task",
  "label":    "legitimate|scam",
  "messages": [
    {"role": "system",    "content": [{"type":"text", "text": "<scam-policy + Philippine context>"}]},
    {"role": "user",      "content": [
        {"type": "image", "image": <PIL.Image>},     # 33-60 frames, RGB
        ...,
        {"type": "text",  "text": "[Audio Transcript]\n... \nTitle: ...\nDescription: ...\n\nReason inside <think>...</think> tags ..."},
    ]},
    {"role": "assistant", "content": [{"type":"text", "text": "<assistant CoT target — see below>"}]},
  ],
}

Cleaning applied

—OCR Temporal Text block stripped — was per-second one-word fragments with frequent misreads, ~600 tokens/sample of pure noise.
—<image> placeholder tokens replaced with PIL Image entries in the user-content list.
—<think> instruction appended to user prompt.
—Assistant target reformatted into V8 Chain-of-Thought-with-think structure (below).

V8 CoT-with-think assistant target format (DeepSeek-R1 style)

<think>
Evidence: <single-line observation derived from frames + audio + title/description>
Criteria hits (C1-C7): <list of Cx hits with reasoning, or "none">
</think>
{"verdict": "Yes", "confidence": 0.95, "category": "Crypto Investment"}

Why this format — research-backed:

—Reasoning genuinely conditions the verdict (real CoT, not post-hoc rationalization) — Anthropic 2023/2025 faithfulness papers, RFEval 2026.
—`</think>` is a hard syntactic boundary — the model learns to emit it reliably, which prevents the rationale-repetition collapse that kills naive verdict-last formats.
—JSON output is bulletproof to parse even if the think block goes off-rails — 4-tier robust parser handles all failure modes.
—Schema matches Qwen3-VL-Thinking's native template — minimal distribution shift on small datasets.

Parsing