README.md

---
dataset_info:
  features:
  - name: messages
    dtype: string
  - name: prompt_length
    dtype: int64
  - name: response_length
    dtype: int64
  splits:
  - name: train
    num_bytes: 423667776
    num_examples: 2000
  download_size: 384009093
  dataset_size: 423667776
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Synthetic Dataset: High Context, Low Generation

## Dataset Description

This is a synthetic benchmark dataset designed to test LLM inference performance in **high-context, low-generation** scenarios. The dataset consists of 2,000 samples with randomly generated tokens that simulate real-world workloads where models process long input contexts but generate relatively short responses.

### Use Cases

This dataset is ideal for benchmarking:
- **Document analysis** with short answers
- **Long-context Q&A** systems
- **Information extraction** from large documents
- **Prompt processing efficiency** and TTFT (Time-to-First-Token) optimization

### Dataset Characteristics

- **Number of Samples**: 2,000
- **Prompt Length Distribution**: Normal distribution
  - Mean: 32,000 tokens
  - Standard deviation: 10,000 tokens
- **Response Length Distribution**: Normal distribution
  - Mean: 200 tokens
  - Standard deviation: 50 tokens
- **Tokenizer**: meta-llama/Llama-3.1-8B-Instruct

### Dataset Structure

Each sample contains:
- `prompt`: A sequence of randomly generated tokens (high context)
- `prompt_length`: Number of tokens in the prompt
- `response_length`: Number of tokens in the response

```python
{
    'prompt': str,
    'prompt_length': int,
    'response_length': int
}
```

### Token Generation

- Tokens are randomly sampled from the vocabulary of the Llama-3.1-8B-Instruct tokenizer
- Each sample is independently generated with lengths drawn from the specified distributions
- The dataset ensures realistic token sequences while maintaining controlled length distributions

## Related Datasets

This dataset is part of a suite of three synthetic benchmark datasets, each designed for different workload patterns:

1. **🔷 synthetic_dataset_high-low** (this dataset)
   - High context (32k tokens), low generation (200 tokens)
   - Focus: Prompt processing efficiency, TTFT optimization

2. **synthetic_dataset_mid-mid**
   - Medium context (1k tokens), medium generation (1k tokens)
   - Focus: Balanced workload, realistic API scenarios

3. **synthetic_dataset_low-mid**
   - Low context (10-120 tokens), medium generation (1.5k tokens)
   - Focus: Generation throughput, creative writing scenarios

## Benchmarking with vLLM

This dataset is designed for use with the vLLM inference framework. The vLLM engine supports a `min_tokens` parameter, allowing you to pass `min_tokens=max_tokens=response_length` for each prompt. This ensures that the response length follows the defined distribution.

### Setup

First, install vLLM and start the server:

```bash
pip install vllm

# Start the vLLM server
vllm serve meta-llama/Llama-3.1-8B-Instruct
```

### Usage Example

```python
from datasets import load_dataset
from openai import OpenAI

# Load the dataset
dataset = load_dataset("jonasluehrs-jaai/synthetic_dataset_high-low")

# Initialize vLLM client (OpenAI-compatible API)
client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="token-abc123",
)

# Use a sample from the dataset
sample = dataset['train'][0]

# Make a completion request with controlled response length
completion = client.chat.completions.create(
    model="meta-llama/Llama-3.1-8B-Instruct",
    messages=[{"role": "user", "content": sample['prompt']}],
    max_tokens=sample['response_length'],
    extra_body={"min_tokens": sample['response_length']},
)

print(f"Generated {len(completion.choices[0].message.content)} characters")
```

For more information, see the [vLLM OpenAI-compatible server documentation](https://docs.vllm.ai/en/v0.8.3/serving/openai_compatible_server.html).

## License

This dataset is released under the MIT License.