Update README.md

README.md

@@ -45,89 +45,113 @@ configs:
         path: synthetic_harmful_prompts_vllm.csv
 ---
 
-# 🛡️ Safety Aligned Datasets
+# Safety Aligned Datasets
 
-> *"You cannot teach a model to refuse what it has never been asked."*
+<div align="center">
 
-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.
+*A high-fidelity adversarial corpus engineered for alignment research, refusal boundary modeling, and robustness evaluation of Small Language Models.*
+
+[![License: CC-BY-4.0](https://img.shields.io/badge/License-CC--BY--4.0-blue.svg)]() [![Split: harmful](https://img.shields.io/badge/Split-harmful-red)]() [![Configs: 6](https://img.shields.io/badge/Configurations-6-orange)]() [![Engine: vLLM](https://img.shields.io/badge/Engine-vLLM-green)]()
+
+</div>
 
 ---
 
-## What's Inside
+## The Problem This Solves
+
+- Fine-tuning a Small Language Model to be safe is not the same as fine-tuning it to *understand* safety.
+
+- Most safety datasets give models clean refusal examples on obvious prompts — and those models fail the moment an adversary wraps a harmful request in a hypothetical, a role-play, or a chain of plausible reasoning. The failure is not in the refusal mechanism. It's in the **coverage gap**: the model has never been exposed to the attack surface it will face in the wild.
+
+- This dataset was built to close that gap.
+
+---
+
+## What Makes It Technically Valuable
+
+### 🎯 Multi-Vector Adversarial Coverage
+
+Unlike single-source safety datasets that skew toward one threat category, this corpus spans **six distinct adversarial configurations** — each targeting a different failure mode in model alignment:
+
+- **`beavertails`** covers the canonical safety meta-categories (violence, hate speech, self-harm) drawn from one of the most widely cited safety benchmarks in the field. It gives your model a principled baseline across well-defined harm taxonomies.
+
+- **`donot_answer`** maps prompts to a structured AI safety evaluation framework, making it ideal for benchmarking against established refusal rubrics and comparing alignment progress across model versions.
+
+- **`forbidden_questions`** is the most adversarially complex configuration — prompts are designed with fictional framing, nested hypotheticals, and philosophical edge cases. This is where naive keyword-based safety mechanisms consistently break down.
 
-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.
+- **`harmfulqa`** focuses on query-style harmful inputs that mimic real user behavior. The risk vectors span multiple domains, making this subset critical for models deployed in open-ended conversational settings.
 
-### ⚠️ 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.
+- **`llm_redteaming`** contains empirical jailbreaks documented during public and academic red-teaming exercises. These are prompts that have *already succeeded* in the wild — your model needs to have seen them.
+
+- **`synthetic_harmful_prompts`** is forward-looking: adversarially generated expansions that simulate novel, unseen attack patterns. This is your zero-day stress test — it pushes models to generalize safe behavior rather than memorize safe responses.
+
+Together, these configurations give a model exposure to the **full adversarial stack**: known attacks, structured taxonomies, behavioral mimicry, fictional obfuscation, and generative novelty.
 
 ---
 
-## Dataset Structure
+### 🧠 Built for SLM Alignment Specifically
 
-One global split — **`harmful`** — partitioned into **6 threat-modeling domains**:
+Large models can absorb safety training with relatively few examples. Small Language Models cannot afford that luxury — they are data-efficient learners with sharper trade-offs between helpfulness and safety.
 
-| Configuration | Source File | What it targets |
-| :--- | :--- | :--- |
-| `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 |
+This dataset is optimized for that constraint:
 
-### Data Fields
+**Supervised Fine-Tuning (SFT) with contrastive pairing** — Each adversarial prompt and its generated response can be paired with a curated safe response to create preference pairs. The model learns not just *what* to refuse, but *how* to refuse in a way that is polite, reasoned, and non-dismissive.
 
-| 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 |
+**Direct Preference Optimization (DPO)** — The `(prompt, safety_res, model)` triplet structure maps naturally onto DPO's chosen/rejected framework. The `safety_res` field provides baseline behavior; researchers can use that as the rejected sample against a human-curated or GPT-4-verified chosen response, enabling efficient preference learning without a reward model.
+
+**False-Refusal Calibration** — Over-refusal is an alignment failure too. Several configurations — particularly `forbidden_questions` and `harmfulqa` — include borderline inputs that use sensitive keywords in genuinely benign contexts. Training on these teaches a model to reason about *intent*, not just surface-level patterns.
 
 ---
 
-## Why This Dataset Exists
+### ⚙️ Generation Pipeline
+
+The dataset was generated using a principled, high-throughput pipeline designed for reproducibility and inference efficiency:
 
-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.
+```
+Base Model  →  TheBloke/Mistral-7B-Instruct-v0.2-AWQ
+Engine      →  vLLM (Vectorized LLM inference framework)
+Hardware    →  NVIDIA T4 GPU · 16GB VRAM · Google Colab
+```
 
-This repository is an empirical sandbox designed to close that gap. Use it to:
+**Why Mistral-7B-Instruct-v0.2-AWQ?** It sits at the intersection of strong instruction-following and accessible compute — making it representative of the exact model class this dataset is designed to align. Responses generated by a model *in the same capability tier* as the target are more realistic training signal than outputs from a frontier model.
+
+**Why vLLM?** Batch processing at scale across six heterogeneous prompt distributions requires a framework that can handle dynamic context lengths without throughput collapse. vLLM's PagedAttention mechanism makes this practical on constrained hardware.
+
+---
 
-- **🎯 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: What Your Model Will Learn to Handle
 
-### Jailbreak Taxonomy
+The adversarial techniques present across configurations cover the full spectrum of known attack classes:
 
-The prompts cover a range of sophisticated attack surfaces:
+**Role-Play Frameworks** — Instructions that force the model into an alternate persona (developer mode, uncensored assistant, fictional AI with no restrictions). These exploit the model's instruction-following strength against its safety constraints.
 
-- **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
+**Hypothetical Contextualization** — Harmful requests buried inside academic thought experiments, historical framing, or fictional narratives. The surface text appears benign; the extracted information is not.
+
+**Linguistic Obfuscation** — Multi-turn logical chains, domain jargon, and structural misdirection designed to bypass keyword filters and confuse intent classifiers. These require the model to reason about the *trajectory* of a conversation, not just the final message.
+
+Exposure to all three classes — individually and in combination — is what separates a model that has learned to refuse from a model that has learned to *stay safe*.
 
 ---
 
-## Generation Pipeline
+## Data Schema
 
-| 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 |
+Each record contains three fields:
 
-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.
+- **`prompt`** — The raw adversarial or harmful input instruction, unmodified from its source
+- **`safety_res`** — The response generated by the inference model during the generation pipeline
+- **`model`** — The specific model variant used, enabling controlled ablation studies across response sources
 
 ---
 
 ## Licensing & Attribution
 
-This dataset is a compilation of open-source data from multiple upstream repositories — each subject to its own license.
-
-| Scope | License |
-| :--- | :--- |
-| Upstream data portions | Original licenses (CC-BY-SA 4.0, MIT, Apache 2.0) |
-| Synthetic responses + framework | CC-BY-4.0 |
+Upstream data portions — drawn from `BeaverTails`, `DoNotAnswer`, `HarmfulQA`, and associated red-teaming sets — retain their original licenses (CC-BY-SA 4.0, MIT, Apache 2.0). Users are responsible for compliance with each upstream source's terms.
 
-Upstream sources include `BeaverTails`, `DoNotAnswer`, `HarmfulQA`, and associated public red-teaming sets. Users are responsible for adhering to their respective terms.
+The unified compilation framework and all synthetic responses are released under **CC-BY-4.0** for research use.
 
 ---
 
-> **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.
+> ⚠️ **Disclaimer:** This dataset contains explicit, toxic, and adversarial content. It exists to advance the science of AI safety, alignment, and defensive model training. The maintainers do not condone or encourage the execution of any harmful behaviors described within these prompts. Deploy responsibly, in controlled research environments only.
+
+
+