Update README.md

README.md

@@ -45,69 +45,89 @@ configs:
         path: synthetic_harmful_prompts_vllm.csv
 ---
 
-# Safety Aligned Datasets
+# 🛡️ Safety Aligned Datasets
 
-## Dataset Summary
-**Safety Aligned Datasets** is a curated collection of adversarial, jailbreak, and policy-violating prompts compiled from various open-source benchmarks and synthetic generation pipelines. Each subset contains the raw `prompt`, the model's generated response (`safety_res`), and the generating `model` identifier. 
+> *"You cannot teach a model to refuse what it has never been asked."*
 
-This dataset is explicitly engineered to support **safety enhancement fine-tuning experiments**, alignment benchmarking, and false-refusal research for Small Language Models (SLMs). 
+A battle-tested corpus of adversarial, jailbreak, and policy-violating prompts — curated from the frontlines of open-source safety research. Built to stress-test alignment, expose refusal boundaries, and make Small Language Models harder to break.
 
-### ⚠️ Critical Usage Notice
-This dataset is strictly intended for **academic research and educational exploration purposes only**. Due to its origin across heterogeneous sources with differing distribution terms, **this dataset is not intended for direct production training or deployment**. Users must handle this data in a controlled sandbox environment.
+---
+
+## What's Inside
+
+Each record is a triplet: the **raw adversarial prompt**, the **model's generated response**, and the **model identifier** that produced it. No sanitization, no softening — this is the unfiltered input the safety community actually needs.
+
+### ⚠️ Usage Notice
+This dataset is strictly for **academic research and educational exploration**. It originates from heterogeneous sources with differing distribution terms and is **not intended for direct production training or deployment**. Handle it inside a controlled sandbox environment.
 
 ---
 
 ## Dataset Structure
 
-The dataset is partitioned under a single global split named **`harmful`**, divided across 6 distinct configurations representing different threat modeling domains:
+One global split — **`harmful`** — partitioned into **6 threat-modeling domains**:
 
-| Configuration Name | Source File | Description |
+| Configuration | Source File | What it targets |
 | :--- | :--- | :--- |
-| `beavertails` | `beavertails_unsafe_vllm.csv` | Prompts targeting core safety meta-categories (violence, hate speech, etc.). |
-| `donot_answer` | `donot_answer_vllm.csv` | High-risk prompts mapped against standard AI safety evaluation frameworks. |
-| `forbidden_questions` | `forbidden_questions_vllm.csv` | Complex/fictional adversarial questions exploring model boundaries. |
-| `harmfulqa` | `harmfulqa_vllm.csv` | Query-based harmful prompts spanning various risk vectors. |
-| `llm_redteaming` | `llm_redteaming_vllm.csv` | Empirical jailbreaks utilized during public and academic red-teaming exercises. |
-| `synthetic_harmful_prompts`| `synthetic_harmful_prompts_vllm.csv` | Generative adversarial expansions to simulate unseen zero-day exploits. |
+| `beavertails` | `beavertails_unsafe_vllm.csv` | Core safety meta-categories: violence, hate speech, and beyond |
+| `donot_answer` | `donot_answer_vllm.csv` | High-risk prompts mapped to standard AI safety evaluation frameworks |
+| `forbidden_questions` | `forbidden_questions_vllm.csv` | Complex, fictional, adversarial queries probing model boundaries |
+| `harmfulqa` | `harmfulqa_vllm.csv` | Query-style harmful prompts spanning a wide range of risk vectors |
+| `llm_redteaming` | `llm_redteaming_vllm.csv` | Empirical jailbreaks from public and academic red-teaming exercises |
+| `synthetic_harmful_prompts` | `synthetic_harmful_prompts_vllm.csv` | Adversarially generated expansions simulating unseen zero-day exploits |
 
 ### Data Fields
-* `prompt` (*string*): The raw adversarial or harmful input instruction.
-* `safety_res` (*string*): The target model output generated during inference.
-* `model` (*string*): The specific inference engine and model variant used.
+
+| Field | Type | Description |
+| :--- | :--- | :--- |
+| `prompt` | `string` | The raw adversarial or harmful input instruction |
+| `safety_res` | `string` | The target model's output during inference |
+| `model` | `string` | The inference engine and model variant used |
 
 ---
 
-## Purpose & Supported Use Cases
+## Why This Dataset Exists
+
+Aligning Small Language Models is hard. You can't fix what you can't measure, and you can't measure what you've never tested.
 
-This repository acts as an empirical sandbox for **alignment engineering for Small Language Models (SLMs)**. It allows safety researchers to:
-* **Supervised Fine-Tuning (SFT) & DPO:** Train SLMs to learn robust, polite, and objective refusal boundaries when paired with contrastive safe data.
-* **Over-Refusal / False-Refusal Analysis:** Study the fine line between true malicious intent and complex, benign queries that use sensitive keywords.
-* **Adversarial Robustness Benchmarking:** Evaluate how easily vanilla or partially aligned models collapse under structured jailbreaks.
+This repository is an empirical sandbox designed to close that gap. Use it to:
 
-### Taxonomy of Jailbreak Vectors Included
-The prompts within these configurations leverage multiple sophisticated adversarial techniques:
-* **Role-Play Frameworks:** Instructions forcing the model into an unaligned persona or developer override simulator.
-* **Hypothetical Contextualization:** Masking harmful requests inside academic, historical, or purely fictional framing.
-* **Linguistic Obfuscation:** Using multi-turn reasoning, complex logical structures, or specialized domain jargon to bypass standard keyword filters.
+- **🎯 SFT & DPO Training** — Pair with contrastive safe data to teach models robust, polite, and objective refusal behavior
+- **🔍 Over-Refusal Analysis** — Study the fine line between genuine malicious intent and benign queries that happen to use sensitive keywords
+- **💥 Adversarial Robustness Benchmarking** — Measure how vanilla or partially aligned models collapse under structured jailbreaks
+
+### Jailbreak Taxonomy
+
+The prompts cover a range of sophisticated attack surfaces:
+
+- **Role-Play Frameworks** — Forcing the model into an unaligned persona or developer-mode override
+- **Hypothetical Contextualization** — Burying harmful requests inside academic, historical, or fictional framing
+- **Linguistic Obfuscation** — Multi-turn reasoning, complex logical structures, or domain jargon designed to slip past keyword filters
 
 ---
 
-## Dataset Generation & Methodology
+## Generation Pipeline
 
-To ensure high throughput and structured generation, the following pipeline was implemented:
+| Component | Detail |
+| :--- | :--- |
+| **Hardware** | NVIDIA T4 GPU (16GB VRAM) via Google Colab |
+| **Base Model** | `TheBloke/Mistral-7B-Instruct-v0.2-AWQ` |
+| **Inference Engine** | vLLM — optimized for batch throughput and context efficiency |
 
-* **Hardware Used:** Google Colab hosted **NVIDIA T4 GPUs** (16GB VRAM).
-* **Base Model:** `TheBloke/Mistral-7B-Instruct-v0.2-AWQ` — chosen for its strong instruction-following capabilities and efficient execution profile.
-* **Inference Engine:** **vLLM** (Vectorized Large Language Model) framework, optimizing batch processing throughput and context efficiency during generation.
+Mistral-7B-Instruct-v0.2-AWQ was chosen for its strong instruction-following and efficient execution profile. vLLM handled the rest, enabling high-throughput structured generation across all configurations.
 
 ---
 
 ## Licensing & Attribution
 
-This dataset is a compilation of data originating from multiple open-source repositories, each subject to their respective upstream licenses (e.g., **CC-BY-SA 4.0, MIT, Apache 2.0**). 
+This dataset is a compilation of open-source data from multiple upstream repositories — each subject to its own license.
 
-* **Upstream Compliance:** The original data portions retain their original licenses. Users are responsible for adhering to the terms set forth by the original authors of `BeaverTails`, `DoNotAnswer`, `HarmfulQA`, and associated red-teaming sets.
-* **Compilation License:** This unified framework and its generated synthetic responses are distributed under a **CC-BY-4.0** license for research purposes.
+| Scope | License |
+| :--- | :--- |
+| Upstream data portions | Original licenses (CC-BY-SA 4.0, MIT, Apache 2.0) |
+| Synthetic responses + framework | CC-BY-4.0 |
+
+Upstream sources include `BeaverTails`, `DoNotAnswer`, `HarmfulQA`, and associated public red-teaming sets. Users are responsible for adhering to their respective terms.
+
+---
 
-### Disclaimer
-> This dataset contains explicit, toxic, and adversarial text inputs that may be offensive or harmful. It is hosted exclusively to advance the science of AI safety, alignment, and defensive model training. The repository maintainers do not condone, support, or encourage the execution of any harmful behaviors described within these prompts.
+> **Disclaimer:** This dataset contains explicit, toxic, and adversarial text that may be offensive or harmful. It is hosted exclusively to advance the science of AI safety, alignment, and defensive model training. The repository maintainers do not condone or encourage the execution of any harmful behaviors described within these prompts.