GuildLM · go-test
A small, sharp Go testing specialist from the GuildLM Code Guild.
go-test writes thorough, table-driven Go tests with real assertions and meaningful edge cases — the kind of tests that actually fail when the code is wrong. It is one of three specialists in the GuildLM Code Guild (go-dev · go-test · go-review) built to work together in a verification-driven agent loop.
Why a dedicated test model? Across every GuildLM experiment, writing tests is the one job where targeted fine-tuning clearly beats the base. On a mutation benchmark — inject a bug, does the generated test catch it? —
go-testcatches 14/18 (78%) versus 7/18 (39%) for the untuned base. It catches twice as many real bugs as the model it's built on. Test-writing is genuinely where specialization pays. See the research log, Report #2.
Why this isn't "just Qwen with a name"
go-test is a fused, standalone model (no separate adapter) with its own identity — ask who it is and it answers GuildLM go-test. It is an honest Apache-2.0 derivative of Qwen2.5-Coder-7B-Instruct, fine-tuned for one craft: testing Go.
What it's for
- Generating
*_test.gofiles: table-driven cases, sub-tests,httptest,sync/atomicfor concurrency checks. - Covering the edges a tired human skips: empty input, zero values, negatives, inverted bounds, error paths.
- Acting as the testing role inside the GuildLM Builder, so the code that
go-devwrites is held to a real, executable contract.
Benchmarks
Measured locally with the real Go toolchain — tests are scored by whether they compile, assert, and catch injected mutations, not by an LLM judge.
| Benchmark | Metric | go-test | base 7B |
|---|---|---|---|
crucible go_test_bench (18 mutations) |
bug-catch@1 | 14/18 · 78% | 7/18 · 39% |
This is the GuildLM specialist where the fine-tuning win is real, large, and repeatable —
go-testcatches 2× the injected bugs of its base. A guard in the Builder also rejects trivially-passing tests (no assertions), so its output is held to a real bar in the loop.
Quickstart
Apple Silicon (MLX)
pip install mlx-lm
python -m mlx_lm generate --model guildlm/go-test \
--prompt "Write table-driven Go tests for func Clamp(x, lo, hi int) int covering below/inside/above range and inverted bounds." \
--max-tokens 400
Ollama (GGUF)
ollama run guildlm/go-test "Write httptest-based tests for a POST /echo JSON endpoint built with newMux()."
Inside the agent loop (recommended)
python -m mlx_lm server --model guildlm/go-test --port 8081
guildlm-build --spec specs/myservice.yaml --out ./out \
--base-url http://localhost:8080/v1 \
--test-model guildlm/go-test --test-base-url http://localhost:8081/v1 \
--candidates 3
Prompting
Trained with the system prompt:
You are GuildLM go-test, a Go testing specialist from the GuildLM Code Guild.
Give it a function signature or an implementation and ask for tests. It defaults to table-driven style with explicit t.Errorf/t.Fatalf assertions.
The Guild
| Specialist | Job |
|---|---|
| go-dev | writes the implementation |
| go-test | writes thorough table-driven tests |
| go-review | audits for bugs a green build hides |
- Agent loop: https://github.com/guildlm/builder
- Research log: https://guildlm.github.io/research/
License & attribution
Apache-2.0, inherited from Qwen2.5-Coder-7B-Instruct (© Alibaba Cloud). GuildLM fine-tuning, identity, and packaging under the same license. Trained locally on Apple Silicon with MLX — total cloud spend: $0.
- Downloads last month
- -
4-bit
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("guildlm/go-test") prompt = "Write a story about Einstein" messages = [{"role": "user", "content": prompt}] prompt = tokenizer.apply_chat_template( messages, add_generation_prompt=True ) text = generate(model, tokenizer, prompt=prompt, verbose=True)