GuildLM · go-lora-adapters

Every LoRA adapter the GuildLM Code Guild trained for Go, in one place, with the number each one scored.

This is the archive behind the three fused specialists (guildlm/go-dev · guildlm/go-test · guildlm/go-review). The fused models are convenient to run; the adapters are what was actually learned, and they are ~100× smaller. Any fused model can be rebuilt from here in one command (see Rebuild a fused model).

Why publish the losers too. GuildLM's bet is that a small specialist wins through the algorithm around it (compile-and-test loop, retrieval, deterministic gates), not through its weights. Measuring that honestly means keeping every adapter that failed to beat its base, with its score. Most of the adapters below are net-negative on the hard unit benchmark. That result is the finding, not a mistake to hide.

The headline measurement (go_dev_bench v2, 48 tasks, greedy, real go build + go test)

Same harness for every row (crucible/mlx_bench.py, direct MLX load, <|im_end|> registered as EOS, generations committed and re-scorable offline with rescore_dev_bench.py).

adapter	recipe	raw	+goimports	vs base (+goimports)
(none — base Qwen2.5-Coder-7B-Instruct-4bit)	—	39	44	—
go-dev-mixed-v4	SFT, mixed data, 1200 iters, r=8	33	40	−4
go-dev-dapt-replay	DAPT (Kaggle) + ~13% chat replay	30	36	−8
go-dev-final	SFT	32	35	−9
go-dev-dapt600	DAPT (Kaggle), 600 steps	29	33	−11
go-dev-mixed-v5	SFT, mixed data v5	29	32	−12
go-dev-dapt300	DAPT (Kaggle), 300 steps	26	31	−13
go-dev-mined	SFT on GitHub-mined data	23	24	−20
(none — base Qwen2.5-Coder-14B-Instruct-4bit)	—	36	43	—
go-dev-14b	SFT on the 14B base	27	40	−3

Reading: every specialist ≤ its base at both 7B and 14B, and most of the deficit is Go hygiene (goimports closes most of it), not reasoning. The one place the adapters earn their keep is as complementary ensemble members: the union of base + all specialists solves 47/48, and adding the 14B members reaches 48/48. No single adapter beats the base. Full log: crucible/RESULT-go-dev-bench-v2.txt in guildlm/guild-code.

Secondary benches (test / review / edit) tell the same story; go-test's apparent +3/18 on the mutation bench turned out to be a validity premium that base + goimports also captures (14/18 vs 11/18). Details: crucible/AUDIT-secondary-benchmarks.txt.

What is in this repo

adapters/<name>/adapter_config.json      exact mlx_lm.lora config (base, data dir, iters, rank, lr)
adapters/<name>/adapters.safetensors     the final adapter
adapters/<name>/NNNNNNN_adapters.safetensors   intermediate checkpoints where they were kept
kaggle-dapt/session1, session2           HF-PEFT checkpoints from the Kaggle DAPT runs (free T4)
kaggle-dapt/replay-session1              the replay-mix DAPT run
kaggle-dapt/ckpt-dataset, replay-dataset the Kaggle dataset bundles those runs were resumed from

Adapter inventory (from each adapter_config.json)

adapter	base	data dir	iters	rank	lr	final weights	kept checkpoints
go-dapt	Qwen2.5-Coder-7B-Instruct-4bit	dapt	6000	8	0.0001	no final	0
go-dapt-core-smoke	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-dapt-core-smoke	600	8	5e-05	yes	3
go-dapt-replay-smoke	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-dapt-replay-smoke	600	8	5e-05	yes	3
go-dev	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data	400	8	0.0001	yes	4
go-dev-0000200	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data	400	8	0.0001	yes	0
go-dev-0000300	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data	400	8	0.0001	yes	0
go-dev-14b	Qwen2.5-Coder-14B-Instruct-4bit	.mlx-data	300	8	8e-05	yes	3
go-dev-15b-mixed-v3	Qwen2.5-Coder-1.5B-Instruct-4bit	.mlx-data-godev-mixed-v3	1200	8	0.0001	yes	6
go-dev-dapt-replay	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt-replay-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	6
go-dev-dapt300	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt-7b-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	6
go-dev-dapt300-ck0000400	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt-7b-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	0
go-dev-dapt300-ck0000800	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt-7b-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	0
go-dev-dapt300-ck0001000	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt-7b-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	0
go-dev-dapt300-ck0001200	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt-7b-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	0
go-dev-dapt600	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt600-7b-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	6
go-dev-dapt600-ck0001000	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt600-7b-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	0
go-dev-dapt600-ck0001200	/Users/fatihturker/Desktop/Personal/Dev/guildlm/.mlx-fused/go-dapt600-7b-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	0
go-dev-final	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-scaled	300	8	8e-05	yes	0
go-dev-id	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-id	200	8	0.0001	yes	0
go-dev-maint	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-maint	160	8	6e-05	yes	0
go-dev-mined	Qwen2.5-Coder-7B-Instruct-4bit	godev2048	4000	8	0.0001	yes	6
go-dev-mixed	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-mixed	1200	8	0.0001	yes	6
go-dev-mixed-v2	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-mixed2	1200	8	0.0001	yes	6
go-dev-mixed-v3	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-mixed-v3-7b	1200	8	0.0001	yes	6
go-dev-mixed-v4	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-mixed-v4-7b	1200	8	0.0001	yes	6
go-dev-mixed-v5	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-mixed-v5-7b	1200	8	0.0001	yes	1
go-dev-mixed-v5b	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-mixed-v5-7b	1000	8	0.0001	yes	5
go-dev-v2	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data	300	8	8e-05	yes	3
go-dev-v2-id	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-v2	260	8	0.0001	yes	0
go-dev-v3-id	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-godev-v2	90	8	4e-05	yes	0
go-review	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-review	180	8	0.0001	yes	2
go-review-id	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-review-id	150	8	0.0001	yes	0
go-review-mined	Qwen2.5-Coder-7B-Instruct-4bit	goreview	1200	8	0.0001	yes	6
go-review-mined-ck0000200	Qwen2.5-Coder-7B-Instruct-4bit	goreview	1200	8	0.0001	yes	0
go-review-mined-ck0000400	Qwen2.5-Coder-7B-Instruct-4bit	goreview	1200	8	0.0001	yes	0
go-review-mined-ck0000600	Qwen2.5-Coder-7B-Instruct-4bit	goreview	1200	8	0.0001	yes	0
go-review-mined-ck0001000	Qwen2.5-Coder-7B-Instruct-4bit	goreview	1200	8	0.0001	yes	0
go-review-mined-ck0001200	Qwen2.5-Coder-7B-Instruct-4bit	goreview	1200	8	0.0001	yes	0
go-review-scaled	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-goreview-scaled	200	8	0.0001	yes	0
go-test	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-test	150	8	0.0001	yes	2
go-test-id	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-test-id	150	8	0.0001	yes	0
go-test-scaled	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-gotest-scaled	220	8	0.0001	yes	0
go-test-v2	Qwen2.5-Coder-7B-Instruct-4bit	.mlx-data-test	120	8	8e-05	yes	2

Use an adapter directly (Apple Silicon, MLX)

pip install mlx-lm huggingface_hub
hf download guildlm/go-lora-adapters --include "adapters/go-dev-mixed-v4/*" --local-dir ./go-lora
python -m mlx_lm.generate \
  --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit \
  --adapter-path ./go-lora/adapters/go-dev-mixed-v4 \
  --prompt "Write an idiomatic Go function that reverses a string by runes."

Rebuild a fused model

python -m mlx_lm.fuse \
  --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit \
  --adapter-path ./go-lora/adapters/go-dev-mixed-v4 \
  --save-path ./go-dev-mixed-v4-fused

The go-dev-14b adapter fuses onto mlx-community/Qwen2.5-Coder-14B-Instruct-4bit and go-dev-15b-mixed-v3 onto mlx-community/Qwen2.5-Coder-1.5B-Instruct-4bit. The *-smoke and *-id adapters are pipeline smoke tests and identity probes and carry no benchmark claim. Each adapter's adapter_config.json names its base; the inventory below is rendered from those files, so it cannot drift from the weights.

Reproduce a score

git clone https://github.com/guildlm/guild-code && cd guild-code/go/crucible
python mlx_bench.py --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit \
  --adapter ./go-lora/adapters/go-dev-mixed-v4 --save-generations out.jsonl
python rescore_dev_bench.py --generations out.jsonl --repair imports

Provenance

• Base models: mlx-community/Qwen2.5-Coder-{1.5B,7B,14B}-Instruct-4bit (Apache-2.0).
• SFT data: compile-verified Go generated with Claude as teacher, plus GitHub-mined Go for the mined and DAPT runs.
• Compute: Apple M1 Max (MLX) for SFT and evaluation; Kaggle free T4 for DAPT. Total spend: $0.
• Every number above was produced by the real Go toolchain, never by an LLM judge.

Part of GuildLM — small, sharp, open specialists, and an honest log of what they can and cannot do.