GuildLM · go-dev

A small, sharp Go development specialist from the GuildLM Code Guild.

go-dev writes clean, idiomatic, standard-library-first Go — types, functions, concurrency, and whole multi-file packages. It is one of three specialists in the GuildLM Code Guild (go-dev · go-test · go-review) designed to be wrapped in a verification-driven agent loop rather than used as a lone chatbot.

The bet: capability = model × algorithm. A 7B specialist inside a compile-and-test loop, grounded by retrieval and guarded by deterministic gates, writes correct backends that a much larger general model — with no scaffolding — does not. go-dev is the model half. The Builder agent loop is the algorithm half.

---

Why this isn't "just Qwen with a name"

go-dev is a fused, standalone model (no separate adapter) with its own identity — ask it who it is and it answers GuildLM go-dev, not the base model. It is fine-tuned for one job (writing Go) and shipped as part of a guild that works together. Under the hood it is an honest Apache-2.0 derivative of Qwen2.5-Coder-7B-Instruct — we attribute the base proudly, and the value we add is specialization + the agent algorithm around it.

What it's for

• Generating idiomatic Go: structs, methods, generics, error handling, concurrency.
• Stdlib-first HTTP services (net/http, ServeMux) — no reflexive third-party routers.
• Working as the implementation role inside the GuildLM Builder: decompose a spec → go-dev writes the code → go-test writes the tests → go build/vet/test → fix → go-review audits.

Benchmarks

Measured locally with the real Go toolchain (no LLM-as-judge). See the research log for the full, honest story — including where fine-tuning helps and where the base and the algorithm are the real levers.

Benchmark	Metric	go-dev	base 7B
crucible go_dev_bench (24 tasks)	pass@1 (real go build+go test)	17/24	19/24
project-level score_backend (in the Builder loop)	build + vet + test	3/3 first try on tractable stdlib specs (numkit, jsonapi, worker-pool)	—

Honest note (this is the whole point of GuildLM): on the solo unit benchmark go-dev lands within measurement noise of its base — for pure code-generation, per-role fine-tuning is not the lever; base choice and the agent loop are. go-dev's real edge shows up at the project level: driven by the Builder with retrieval grounding, it writes whole stdlib backends that build, vet and test green on the first try (score_backend 3/3) — which a lone model, prompted once, does not. Use it in the loop; that's where it shines.

Quickstart

Apple Silicon (MLX)

pip install mlx-lm
python -m mlx_lm generate --model guildlm/go-dev \
  --prompt "Write an idiomatic Go function MergeIntervals(intervals [][]int) [][]int." \
  --max-tokens 400

Ollama (GGUF)

ollama run guildlm/go-dev "Write a stdlib-only Go net/http key/value service with GET/PUT."

Inside the agent loop (recommended)

# serve OpenAI-compatible, then let the Builder drive it
python -m mlx_lm server --model guildlm/go-dev --port 8080
guildlm-build --spec specs/myservice.yaml --out ./out \
  --base-url http://localhost:8080/v1 \
  --test-model guildlm/go-test --review-model guildlm/go-review \
  --examples examples/verified_contracts.jsonl --candidates 3

Prompting

go-dev is trained with the system prompt:

You are GuildLM go-dev, a Go development specialist from the GuildLM Code Guild.

Ask for complete, runnable Go. It prefers the standard library and will avoid third-party dependencies unless you explicitly ask.

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 (every experiment, wins and losses): https://guildlm.github.io/research/

License & attribution

Apache-2.0, inherited from the base model Qwen2.5-Coder-7B-Instruct (© Alibaba Cloud). GuildLM fine-tuning, identity, packaging, and the agent loop are released under the same license. All training was done locally on Apple Silicon with MLX — total cloud spend: $0.