VijayRam1812/safety_dataset

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	5256
Assistant response only	5256
Mixed Conversations	5256

Label 0 (Benign): 15768 total

• CSAE: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• SHS: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• VC: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• IP: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• PII: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• DEF: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• SCAM: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• ESP: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• CBRN: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}

Label 1 (Harmful): 15768 total

• CSAE: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• SHS: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• VC: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• IP: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• PII: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• DEF: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• SCAM: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• ESP: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}
• CBRN: 1752 -> {'user_only': 584, 'single_turn': 584, 'multi_turn': 584}

Total samples: 31536 (expected 2 x 9 x 3 x 584 = 31536)

Data Curation

Step 1 — Ingest & Compile existing open-source datasets into a unified {messages, label, source} format:

• BeaverTails (PKU-Alignment) — single-turn prompt/response pairs with is_safe labels
• PKU-SafeRLHF — single-turn with two responses per prompt, each independently labeled safe/unsafe
• Anthropic HH-RLHF — multi-turn Human/Assistant conversations; chosen mapped to benign, rejected to harmful; parsed from Anthropic's custom text format into [{role, content}]

Step 2 — Conversation Type Mutation. Each sample is assigned a conv_type:

• Messages with > 2 turns become multi_turn
• Of the remaining single-turn pairs, ~20% are randomly mutated to user_only by dropping the assistant response
• The rest stay single_turn
• A raw_text column is built: "USER: ...\nASSISTANT: ..."

Step 3 — vLLM Category Classification. Qwen 2.5-7B-Instruct-AWQ classifies each sample into one of the 9 safety categories:

• A zero-shot system prompt lists all 9 category abbreviations with descriptions
• The model reads raw_text and returns the best-matching abbreviation
• Unclassified samples are filtered out
• Capped at 30,000 samples for GPU efficiency

Step 4 — Hierarchical 3D Balancing. A two-pass algorithm balances across all three dimensions:

• Computes target counts: target_total / 2 labels / 9 categories / 3 types
• Pass 1: For each (label, category, type) cell, take up to target_per_type samples
• Pass 2: Any shortfall in a type is backfilled from leftover samples in the same (label, category) group

Step 5 — Gap Analysis & Generation-Based Filling. After balancing, the analyze_and_standardize function:

1. Analyzes every (label, category, conv_type) cell and prints a BEFORE summary
2. Computes the target per cell (max count across all 54 cells)
3. Generates new samples for under-filled cells using vLLM (Qwen 2.5-7B-Instruct-AWQ):
   • For each (category, conv_type) with shortfall, builds a contrastive generation prompt that produces a JSON array of 2 items (1 harmful + 1 benign)
   • The prompt is tailored per conv_type: user_only requires exactly 1 message, single_turn requires user + assistant, multi_turn requires 4-6 alternating messages
   • Generated messages are normalized (handles :role/-role key quirks) and raw_text is populated
4. Standardizes — any residual shortfall is filled by sampling with replacement; excess cells are trimmed down
5. Prints an AFTER summary with final counts

Step 6 — Gemma Formatting & Stratified Split:

• Messages are normalized for Gemma's chat template (alternating user/assistant, merges consecutive same-role)
• label is cast to ClassLabel(["benign", "harmful"])
• A composite stratify_key = "{label}_{category}" enables stratified train/val split (90/10) that preserves proportions
• Final DatasetDict with train and val splits is saved and pushed to the Hub

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