Nerdsking HumanEval Benchmark for llama.cpp (GGUF)

A strict, auditable HumanEval benchmark runner for GGUF models served via llama.cpp, using its OpenAI-compatible HTTP API.

This project focuses on correct execution semantics and reproducibility:

• Prompts are preserved verbatim (no stripping or truncation).
• Only fenced Python code is accepted.
• Each task is executed using strict HumanEval semantics.
• Full outputs and failure reasons are saved for auditing.

Key Features

✅ Correct HumanEval Semantics (Strict)

For every task, execution follows exactly:

1. Execute the original prompt (function signature + docstring)
2. Execute the model-generated code
3. Execute the test harness
4. Call check(entry_point)

✅ Prompt Integrity

• HumanEval prompts are used verbatim
• No stripping, rewriting, or truncation
• Only a minimal instruction header is prepended
• Raw prompts are stored in the output JSON

✅ Strict Code Extraction

Only code inside a single fenced block is accepted:

```python
# code here

If no such block exists → **automatic failure (`no_code`)**.

---

### ✅ Full Failure Attribution
Each failed task records:
- `error_type`
- `error_detail`
- `full_response`
- `generated_code`

---

### ✅ llama.cpp Native Support
- Automatic server start/stop **or** reuse an existing server
- Uses `/v1/completions` OpenAI-compatible API
- Streaming-safe and timeout-safe
- **GGUF-only by design**

---

## Repository Structure

. ├── benchmark.py ├── HumanEval.jsonl ├── LICENSE ├── README.md └── eval_utils/ ├── init.py ├── bench_config.json └── code_bench.py

---

## Dependencies

### Required
- Python 3.10+
- llama.cpp (with server support)
- GGUF model

### Python packages
```bash
pip install requests datasets

Building llama.cpp

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
mkdir -p build && cd build
cmake .. -DLLAMA_CUBLAS=ON
cmake --build . --config Release

Usage

Automatic server management

python benchmark.py --model model.gguf --server-path /path/to/llama.cpp/build/bin

Use existing server

python benchmark.py --server-url http://127.0.0.1:8080 --no-server

Use local HumanEval

python benchmark.py --humaneval-jsonl HumanEval.jsonl

Output

A JSON file is generated containing:

• Full configuration
• Per-task results
• Raw model outputs
• Error attribution
• Timing metrics

License / citation

-If you use this code in research or benchmarking, please cite:

https://github.com/nerdskingcom/gguf_benchmark, IMNECHO / Nerdsking.com