Skip to content

Personal AI, On Personal Devices

Python
1.0 KB
Raw Blame
[recipe]
name = "coding-benchmark"
kind = "discrete"
description = "ReAct coding agent benchmarked against TerminalBench and SWE-bench"
version = "0.1.0"

[intelligence]
model = "qwen3:8b"
quantization = "q4_K_M"

[engine]
key = "ollama"

[agent]
type = "native_react"
max_turns = 20
temperature = 0.2
tools = ["file_read", "file_write", "shell_exec", "code_interpreter", "apply_patch", "think"]
system_prompt = """You are an expert programmer and systems engineer. You solve coding tasks by reasoning through problems step-by-step, reading code, writing solutions, and verifying them with tests.

Approach each task methodically:
1. Understand what is being asked
2. Explore relevant files and code
3. Plan your solution
4. Implement changes carefully
5. Test and verify your work

Use the tools available to read files, write code, execute commands, and apply patches. Always verify that your changes work correctly before concluding."""

[learning]
routing = "heuristic"

[eval]
benchmarks = ["terminalbench", "swebench"]
backend = "jarvis-agent"
max_samples = 50