docs: add airtbench dataset readme

dataset/README.md

@@ -1,2 +1,155 @@
-# AIRTBench Dataset
+# AIRTBench Dataset - External Release
 
+- [AIRTBench Dataset - External Release](#airtbench-dataset---external-release)
+  - [Overview](#overview)
+  - [Dataset Statistics](#dataset-statistics)
+  - [Model Success Rates](#model-success-rates)
+  - [Challenge Difficulty Distribution](#challenge-difficulty-distribution)
+  - [Data Dictionary](#data-dictionary)
+    - [Identifiers](#identifiers)
+    - [Primary Outcomes](#primary-outcomes)
+    - [Performance Metrics](#performance-metrics)
+    - [Resource Usage](#resource-usage)
+    - [Cost Analysis](#cost-analysis)
+    - [Conversation Content](#conversation-content)
+    - [Error Analysis](#error-analysis)
+  - [Usage Examples](#usage-examples)
+    - [Basic Analysis](#basic-analysis)
+    - [Cost Analysis](#cost-analysis-1)
+    - [Performance Analysis](#performance-analysis)
+    - [Conversation Content](#conversation-content-1)
+  - [Contact](#contact)
+  - [Version History](#version-history)
+
+## Overview
+
+This dataset contains the complete experimental results from the AIRTBench paper: "*AIRTBench: An AI Red Teaming Benchmark for Evaluating Language Models' Ability to Autonomously Discover and Exploit AI/ML Security Vulnerabilities.*"
+
+The dataset includes 8,066 experimental runs across 12 different language models and 70 security challenges and is available [here](https://huggingface.co/datasets/dreadnode/AIRTBench/).
+
+## Dataset Statistics
+
+- **Total Runs**: 8,066
+- **Unique Models**: 12
+- **Unique Challenges**: 70
+- **Success Rate**: 20.5%
+
+## Model Success Rates
+
+| Model | Success Rate |
+|---|---|
+| claude-3-7-sonnet-20250219 | 46.86% |
+| gpt-4.5-preview | 36.89% |
+| gemini/gemini-2.5-pro-preview-05-06 | 34.29% |
+| openai/o3-mini | 28.43% |
+| together_ai/deepseek-ai/DeepSeek-R1 | 26.86% |
+| gemini/gemini-2.5-flash-preview-04-17 | 26.43% |
+| openai/gpt-4o | 20.29% |
+| gemini/gemini-2.0-flash | 16.86% |
+| gemini/gemini-1.5-pro | 15.14% |
+| groq/meta-llama/llama-4-scout-17b-16e-instruct | 1.00% |
+| groq/qwen-qwq-32b | 0.57% |
+| groq/llama-3.3-70b-versatile | 0.00% |
+
+## Challenge Difficulty Distribution
+
+| Difficulty | Count |
+|---|---|
+| easy | 4,259 |
+| medium | 2,657 |
+| hard | 1,150 |
+
+## Data Dictionary
+
+### Identifiers
+- **model**: Original model name from API
+- **model_family**: Model provider/family (Anthropic, OpenAI, Google, etc.)
+- **challenge_name**: Name of the security challenge
+- **challenge_difficulty**: Difficulty level (Easy/Medium/Hard)
+
+### Primary Outcomes
+- **flag_found**: Boolean indicating if the run found the flag.
+
+### Performance Metrics
+- **total_flag_submissions**: Total number of flag submissions attempted
+- **correct_flag_submissions**: Number of correct flag submissions (led to success)
+- **incorrect_flag_submissions**: Number of incorrect flag submissions (failed)
+- **duration_minutes**: Total runtime in minutes
+
+### Resource Usage
+- **input_tokens**: Number of input tokens consumed (integer)
+- **output_tokens**: Number of output tokens generated (integer)
+- **total_tokens**: Total tokens (input + output) (integer)
+- **execution_spans**: Number of execution spans
+
+### Cost Analysis
+- **total_cost_usd**: Total cost in USD for the run
+- **input_cost_usd**: Cost for input tokens in USD
+- **output_cost_usd**: Cost for output tokens in USD
+- **tokens_per_dollar**: Number of tokens per dollar spent
+
+### Conversation Content
+- **conversation**: Complete conversation including all chat messages (API keys redacted)
+
+### Error Analysis
+- **hit_rate_limit**: Boolean indicating if rate limits were hit
+- **rate_limit_count**: Number of rate limit errors encountered
+
+## Usage Examples
+
+### Basic Analysis
+```python
+import pandas as pd
+
+# Load the dataset
+df = pd.read_parquet('airtbench_external_dataset.parquet')
+
+# Calculate success rates by model
+success_by_model = df.groupby('model')['flag_found'].mean().sort_values(ascending=False)
+print(success_by_model)
+
+# Calculate success rates by challenge
+success_by_challenge = df.groupby('challenge_name')['flag_found'].mean().sort_values(ascending=False)
+print(success_by_challenge)
+```
+
+### Cost Analysis
+```python
+# Analyze cost efficiency
+cost_analysis = df.groupby('model').agg({
+    'total_cost_usd': 'mean',
+    'cost_per_success': 'mean',
+    'tokens_per_dollar': 'mean',
+    'flag_found': 'mean'
+}).round(4)
+print(cost_analysis)
+```
+
+### Performance Analysis
+```python
+# Analyze performance metrics
+performance = df.groupby('model').agg({
+    'duration_minutes': 'mean',
+    'total_tokens': 'mean',
+    'execution_spans': 'mean',
+    'flag_found': 'mean'
+}).round(2)
+print(performance)
+```
+
+### Conversation Content
+
+```python
+# Example of conversation content
+conversation = df.loc[50, 'conversation']
+conversation = eval(conversation) # Convert string to list
+print(conversation)
+```
+
+## Contact
+
+For questions about this dataset, please contact [[email protected]](mailto:[email protected]).
+
+## Version History
+
+- v1.0: Initial external release