Upload README.md with huggingface_hub

README.md

@@ -1,50 +1,40 @@
 ---
 license: mit
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: test
-    path: data/test-*
-dataset_info:
-  features:
-  - name: conversation
-    list:
-    - name: content
-      dtype: string
-    - name: role
-      dtype: string
-  - name: origin
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 35494601343
-    num_examples: 4074222
-  - name: test
-    num_bytes: 358533438
-    num_examples: 41154
-  download_size: 35787321887
-  dataset_size: 35853134781
 ---
 # Guard-GLP Benign Conversations
 
 Sanitized collection of benign multi-turn conversations, using **train splits only**.
 
+## Pipeline
+1. Load LMSYS Chat 1M and WildChat-4.8M (sharded across 8 worker(s))
+2. Filter harmful/adversarial samples via OpenAI moderation labels
+3. Normalise conversation format across sources
+4. Exact de-duplication (conversation hash)
+6. De-contaminate against WildJailbreak (train + eval configs)
+7. Hold out a uniformly sampled test split
+
 ## Composition
 | Source | Samples |
 |--------|---------|
-| lmsys | 942,157 |
-| wildchat | 524,274 |
-| wildguard | 7,778 |
-| wildjailbreak | 128,781 |
-| **Total** | **1,602,990** |
+| lmsys | 941,476 |
+| wildchat_4m | 3,173,900 |
+| **Total** | **4,115,376** |
+
+## Splits
+| Split | Samples |
+|-------|---------|
+| train | 4,074,222 |
+| test | 41,154 |
 
-**Total tokens:** 1,124,232,118 (Llama 3.2 1B Instruct tokenizer)
+Test split is a uniformly sampled 1% holdout (seed 42).
+
+Total tokens (meta-llama/Meta-Llama-3.1-8B-Instruct chat template): **10,302,302,353**
 
 ## Sanitization
-- **LMSYS Chat 1M & WildChat**: filtered via OpenAI moderation labels (all flagged categories removed)
+- **LMSYS Chat 1M, WildChat & WildChat-4.8M**: filtered via OpenAI moderation labels (all flagged categories removed)
 - **WildGuardMix**: removed adversarial and harmful-labeled prompts/responses
-- **WildJailbreak**: removed entries with harmful data types
+- Exact de-duplication across all sources
+
+- Embedding-based de-contamination against WildJailbreak (Qwen3-Embedding-8B, cosine similarity > 0.99 removed)
 
 Only benign conversations are retained.