Update README.md

README.md

@@ -1,48 +1,46 @@
 ---
-annotations_creators:
-  - no-annotation
-language:
-  - en
-language_creators:
-  - machine-generated
-multilinguality:
-  - monolingual
-size_categories:
-  - 10K<n<100K
-source_datasets:
-  - original
+license: cc-by-4.0
 task_categories:
-  - text-generation
+- text-generation
 task_ids:
-  - text2text-generation
-license: cc-by-4.0
-pretty_name: Safety Aligned Datasets
-viewer: true
+- text2text-generation
+language:
+- en
+size_categories:
+- 10K<n<100K
 configs:
-  - config_name: beavertails
-    data_files:
-      - split: harmful
-        path: beavertails_unsafe_vllm.csv
-  - config_name: donot_answer
-    data_files:
-      - split: harmful
-        path: donot_answer_vllm.csv
-  - config_name: forbidden_questions
-    data_files:
-      - split: harmful
-        path: forbidden_questions_vllm.csv
-  - config_name: harmfulqa
-    data_files:
-      - split: harmful
-        path: harmfulqa_vllm.csv
-  - config_name: llm_redteaming
-    data_files:
-      - split: harmful
-        path: llm_redteaming_vllm.csv
-  - config_name: synthetic_harmful_prompts
-    data_files:
-      - split: harmful
-        path: synthetic_harmful_prompts_vllm.csv
+- config_name: beavertails
+  data_files:
+  - split: harmful
+    path: beavertails_unsafe_vllm.csv
+- config_name: donot_answer
+  data_files:
+  - split: harmful
+    path: donot_answer_vllm.csv
+- config_name: forbidden_questions
+  data_files:
+  - split: harmful
+    path: forbidden_questions_vllm.csv
+- config_name: harmfulqa
+  data_files:
+  - split: harmful
+    path: harmfulqa_vllm.csv
+- config_name: llm_redteaming
+  data_files:
+  - split: harmful
+    path: llm_redteaming_vllm.csv
+- config_name: synthetic_harmful_prompts
+  data_files:
+  - split: harmful
+    path: synthetic_harmful_prompts_vllm.csv
+- config_name: prompt_injection_unsafe_vllm
+  data_files:
+  - split: harmful
+    path: prompt_injection_unsafe_vllm.csv
+- config_name: prompt_injection_safe_vllm
+  data_files:
+  - split: safe
+    path: prompt_injection_safe_vllm.csv
 ---
 
 # Safety Aligned Datasets
@@ -51,7 +49,7 @@ configs:
 
 *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)]()
+[![License: CC-BY-4.0](https://img.shields.io/badge/License-CC--BY--4.0-blue.svg)]() [![Harmful Split](https://img.shields.io/badge/Split-harmful-red)]() [![Safe Split](https://img.shields.io/badge/Split-safe-brightgreen)]() [![Configs: 8](https://img.shields.io/badge/Configurations-8-orange)]() [![Engine: vLLM](https://img.shields.io/badge/Engine-vLLM-green)]()
 
 </div>
 
@@ -59,91 +57,117 @@ configs:
 
 ## The Problem This Solves
 
-- Fine-tuning a Small Language Model to be safe is not the same as fine-tuning it to *understand* safety.
+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.
+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 is in the **coverage gap**: the model has never been exposed to the attack surface it will actually face in deployment.
 
-- This dataset was built to close that gap.
+This dataset was built to close that gap — and to go further. Because alignment is not just about refusal. It is about teaching a model to handle adversarial manipulation at the system level, including direct attacks on its own instruction context.
 
 ---
 
-## What Makes It Technically Valuable
+## Dataset Overview
+
+This corpus spans **eight configurations** across **two splits**, each targeting a distinct failure mode in model alignment.
 
-### 🎯 Multi-Vector Adversarial Coverage
+### The `harmful` Split
 
-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:
+This split contains adversarial prompts and model-generated responses representing unsafe behavior across seven configurations. Each configuration covers a different attack class or harm domain.
 
-- **`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.
+**`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. 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.
+**`donot_answer`**
+Prompts mapped to a structured AI safety evaluation framework. Best suited for benchmarking against established refusal rubrics and tracking 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.
+**`forbidden_questions`**
+The most adversarially complex configuration. Prompts rely on fictional framing, nested hypotheticals, and philosophical edge cases — exactly where naive keyword-based safety mechanisms consistently break down.
 
-- **`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.
+**`harmfulqa`**
+Query-style harmful inputs that closely mimic real user behavior. Risk vectors span multiple domains, making this subset particularly important for models deployed in open-ended conversational settings.
 
-- **`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.
+**`llm_redteaming`**
+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 before deployment.
 
-- **`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.
+**`synthetic_harmful_prompts`**
+Adversarially generated expansions that simulate novel, unseen attack patterns. A zero-day stress test designed to push models toward generalizing safe behavior rather than memorizing 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.
+**`prompt_injection_unsafe_vllm`**
+Targets a distinct and increasingly critical threat class: prompt injection attacks. These prompts attempt to override or hijack the model's system-level instructions through crafted user inputs — a common real-world attack vector in agentic and tool-augmented LLM deployments. Exposure to this subset is essential for any model operating in a pipeline where user input reaches the inference context alongside a system prompt.
 
 ---
 
-### 🧠 Built for SLM Alignment Specifically
+### The `safe` Split
 
-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.
+**`prompt_injection_safe_vllm`**
+The structural counterpart to the unsafe injection configuration above. These records show the model correctly maintaining its intended behavior under injection pressure — ignoring override instructions and staying within its sanctioned boundaries. When paired with the unsafe configuration in a DPO or contrastive SFT setup, these serve as the *chosen* side of the preference pair.
 
-This dataset is optimized for that constraint:
+---
 
-**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.
+## How to Use This Dataset
 
-**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.
+### Supervised Fine-Tuning (SFT)
 
-**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.
+Each adversarial prompt and its generated response can be paired with a curated safe response to create contrastive training examples. The model learns not just what to refuse, but how to refuse in a way that is polite, reasoned, and non-dismissive.
 
----
+### Direct Preference Optimization (DPO)
 
-### ⚙️ Generation Pipeline
+The `(prompt, safety_res, model)` triplet structure maps naturally onto DPO's chosen/rejected framework. The `safety_res` field provides the baseline behavior to use as the rejected sample, paired against a human-curated or verified chosen response. For the prompt injection configurations specifically, `prompt_injection_unsafe_vllm` and `prompt_injection_safe_vllm` form a ready-made contrastive pair without any additional annotation overhead.
 
-The dataset was generated using a principled, high-throughput pipeline designed for reproducibility and inference efficiency:
+### False-Refusal Calibration
 
-```
-Base Model  →  TheBloke/Mistral-7B-Instruct-v0.2-AWQ
-Engine      →  vLLM (Vectorized LLM inference framework)
-Hardware    →  NVIDIA T4 GPU · 16GB VRAM · Google Colab
-```
+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 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.
+### Prompt Injection Robustness
 
-**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.
+With LLMs increasingly deployed in agentic workflows, RAG pipelines, and tool-use environments, resistance to instruction hijacking is no longer optional. The injection split pair directly addresses this training gap, which most general-purpose safety datasets ignore entirely.
 
 ---
 
-### 🗂️ Jailbreak Taxonomy: What Your Model Will Learn to Handle
+## Adversarial Technique Reference
 
-The adversarial techniques present across configurations cover the full spectrum of known attack classes:
+The prompts across all configurations cover four primary attack classes:
 
-**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**
+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 directly against its safety constraints.
 
-**Hypothetical Contextualization** — Harmful requests buried inside academic thought experiments, historical framing, or fictional narratives. The surface text appears benign; the extracted information is not.
+**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.
+**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 in isolation.
 
-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*.
+**Prompt Injection**
+Direct attempts to override system-level instructions through adversarially crafted user inputs. Variants include goal hijacking (redirecting the model to a different task), prompt leakage (extracting the system prompt), and jailbreak injection (embedding role-play attacks inside what appears to be ordinary user data). Particularly relevant for RAG systems, tool-augmented agents, and any deployment where user input coexists with a privileged system context.
+
+---
+
+## Generation Pipeline
+
+```
+Base Model  →  TheBloke/Mistral-7B-Instruct-v0.2-AWQ
+Engine      →  vLLM (Vectorized LLM inference framework)
+Hardware    →  NVIDIA T4 GPU · 16GB VRAM · Google Colab
+```
+
+**Why Mistral-7B-Instruct-v0.2-AWQ?**
+It sits at the intersection of strong instruction-following capability and accessible compute — representative of the exact model class this dataset is designed to align. Responses generated by a model in the same capability tier as the training target produce more realistic and transferable training signal than outputs from a frontier model.
+
+**Why vLLM?**
+Batch processing across eight heterogeneous prompt distributions requires a framework that handles dynamic context lengths without throughput collapse. vLLM's PagedAttention mechanism makes this practical on constrained hardware.
 
 ---
 
 ## Data Schema
 
-Each record contains three fields:
+Each record across all configurations contains three fields:
 
 - **`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
+- **`safety_res`** — The response generated by the inference model during the pipeline run
 - **`model`** — The specific model variant used, enabling controlled ablation studies across response sources
 
 ---
 
-## Licensing & Attribution
+## Licensing and Attribution
 
 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.
 
@@ -151,7 +175,4 @@ The unified compilation framework and all synthetic responses are released under
 
 ---
 
-> ⚠️ **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.
-
-
-
+> **Disclaimer:** This dataset contains explicit, toxic, and adversarial content. It is hosted exclusively 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. Use only in controlled research environments.