VijayRam1812/safety_dataset_synthetic_eval

Dataset Summary:

Dimensions

• Labels (2): 0 = Benign, 1 = Harmful
• Categories (9): CSAE, SHS, VC, IP, PII, DEF, SCAM, ESP, CBRN
• Conversation Types (3): user_only, single_turn, multi_turn
• Target cells: 2 x 9 x 3 = 54 unique combinations

Harmful Content Categories (9 categories):

• Child Sexual Abuse and Exploitation and Sex Crimes
• Self-Harm and Suicide
• Illegal Activities and Violent Crimes
• Intellectual Property or Copyright Violations
• Privacy or PII Violations
• Defamation, Libel, or Slander
• Defrauding, Scamming, Spamming, or Phishing
• Espionage, Spying, Stalking, Hacking, or Doxing
• Chemical, Biological, Radiological, and Nuclear (CBRN) Threats

Benign Content Categories (9 categories):

• Child Sexual Abuse and Exploitation and Sex Crimes
• Self-Harm and Suicide
• Illegal Activities and Violent Crimes
• Intellectual Property or Copyright Violations
• Privacy or PII Violations
• Defamation, Libel, or Slander
• Defrauding, Scamming, Spamming, or Phishing
• Espionage, Spying, Stalking, Hacking, or Doxing
• Chemical, Biological, Radiological, and Nuclear (CBRN) Threats

Sources

• https://huggingface.co/datasets/PKU-Alignment/BeaverTails
• https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF
• https://huggingface.co/datasets/Anthropic/hh-rlhf
• Synthetic dataset generated using Qwen model and vLLM

Conversation type	Samples
User query only	630
Assistant response only	630
Mixed Conversations	630

Label 0 (Benign): 1890 total

• CSAE: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• SHS: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• VC: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• IP: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• PII: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• DEF: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• SCAM: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• ESP: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• CBRN: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}

Label 1 (Harmful): 1890 total

• CSAE: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• SHS: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• VC: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• IP: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• PII: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• DEF: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• SCAM: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• ESP: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}
• CBRN: 210 -> {'user_only': 70, 'single_turn': 70, 'multi_turn': 70}

Total samples: 3780 (expected 2 x 9 x 3 x 70 = 3780)

Data Curation

Step 1 — Bulk Synthetic Generation. No existing datasets — everything is generated from scratch:

• 70 requests per (category, conv_type) combo = 70 x 9 x 3 = 1,890 vLLM calls
• Each call uses the same contrastive prompt as Track 1, producing 2 samples (1 harmful + 1 benign)
• Target: ~3,780 synthetic samples

Step 2 — Gap Analysis & Standardization. Identical to Track 1's Step 5:

• Analyzes the 54 cells for imbalance (JSON parse failures cause uneven counts)
• Generates additional samples for under-filled cells
• Standardizes all cells to the same count

Step 3 — Gemma Normalization:

• Messages normalized for Gemma alternation, raw_text built
• Saved to disk and pushed to the Hub as the eval dataset

---

Key Design Decisions

Aspect	Choice
Generation model	Qwen 2.5-7B-Instruct-AWQ (fast, fits Colab T4)
Target model	Gemma 3-1B-IT (lightweight classifier to be trained)
Prompt strategy	Contrastive pairs — each call produces both a harmful and benign example on the same topic, ensuring the model learns the boundary
Balancing	Two-stage: coarse hierarchical balancing first, then fine-grained per-cell standardization with generation to minimize duplicate sampling
Message normalization	Handles Hub dataset quirks (:role, -role keys) and enforces alternating user/assistant for Gemma compatibility