Teaching a local coding agent from its own mistakes: DPO on a 30B model

The WUIC Assistant is our agentic VS Code plugin: it scaffolds Angular components, dashboards, reports, workflows, and metadata patches for apps built on the WUIC framework. Under the hood it runs qwen3-coder:30b — a Mixture-of-Experts model with ~3B active parameters — served locally through Ollama on a single RTX 4090. No API calls, no data leaving the machine.

For weeks the plugin had been sitting at 100% on its curriculum: 32 tasks, five green runs each, a 96/96 regression suite. Beautiful numbers, and a problem. A curriculum that always passes has stopped measuring anything. It could no longer tell us whether the model generalized to the prompts a real developer types, nor how often it took a wrong turn before correcting itself. We wanted two things:

STOP: use scaffold, not ng generate

), and only This post is the story of how we got there. It involves five OutOfMemory crashes, a saturated benchmark, two hidden contaminations, and a 30-billion-parameter model that did learn — though not exactly what we asked it to.

A model that evaluates itself on the data it was aligned to is lying to itself. So we built eval-only sets, with one iron rule: never in training, never curricularized. If you stabilize a task by adding a dedicated mechanism, that task loses its value as a measurement — and it gets replaced.

The latest of these, the fourth "HARD" set (Q01–Q12), deliberately raises the bar on the levers where the base model actually struggles:

Measured on a clean bench, the base model lands at 25/36 (69%) — failing Q02, Q04 and Q07 on every attempt, and Q05 on two of three. But the most interesting data point was elsewhere: Q01 passed 3 times out of 3 — with 9 redirects. It got there, but by bouncing off the guards the whole way. That's the churn we wanted gone.

Before touching the model, a deceptively innocent question — "is the test environment actually clean and in sync?" — opened a can of worms.

Contamination #1: the metadata reset. Our between-runs reset ran DELETE WHERE id > baseline

— it deleted the rows that had been added, but never reverted the UPDATEs (like hide_in_list

) applied to pre-existing rows. Run after run, columns on the customers

table were progressively hidden, until route_columns

started answering "no columns found." Worse: some of our test prompts had been reworded over time to compensate for this degradation — meaning we'd been fixing the symptoms of an environment bug and mistaking it for backend behavior. The fix: restore the full column set from the clean baseline (metadataTutorial

) on every reset. The general lesson we kept: a reset has to cover what you mutate, not just what you add.

Contamination #2: the report generator. A shared file (gen-report-from-template.mjs ) had been overwritten by a model run, shrinking from 246 lines to 60 — and our reset didn't cover that directory. Restored, added to the reset scope, and write-guarded.

The moral: before you trust the numbers, trust the environment. A test bench that drifts slowly produces phantom regressions that look exactly like model bugs.

With a clean bench, the third holdout set went from 22/36 to 36/36. The fixes were generic — synonyms in the detectors, correct precedence between task "kinds", scaffold guards for reports — never a patch for an individual prompt. That distinction is the whole game: you generalize the root condition, you don't enumerate the cases.

But the churn remained. And churn isn't solved by adding one more guard — it's solved by teaching the model not to fumble the first move. That meant touching the model itself.

The right technique here isn't more supervised fine-tuning. The SFT we'd already tried masked the rejected turns — the model never got the signal "this move is wrong." Redirects dropped only from 43 to 39.

DPO (Direct Preference Optimization) learns from the pair instead: for the exact same context, prefer the accepted action over the rejected one. It's a perfect fit for our problem, because every redirect in the historical transcripts is already a ready-made pair:

ng generate component…

→ STOP) = scaffold widget=map → Saved) = An extractor pulled 559 pairs from more than a thousand archived runs (including the red ones — a pair is valuable if it contains a valid rejected→corrected transition), deduplicated and balanced so no single task dominates.

Here's where theory meets 80 GB of VRAM. DPO does four forward passes per step (policy and reference, over chosen and rejected). And our system prompt — the .clinerules

that give the model all the WUIC context — is enormous: ~20,000 tokens. Multiplied by four passes, the 30B model went OOM before step 0, five times in a row, first on a 4090 under WSL2 and then even on an 80 GB A100.

The debug path, condensed:

peft

.[seq × vocab≈150k] → OOM.lm_head

• loss without ever materializing the full logits) promised to solve it… but on a 4-bit quantized PeftModel

it was SMOKE_FAIL

with no clear traceback.The breakthrough was a counterintuitive one: the real problem wasn't DPO, it was the system prompt. We trimmed the system block in the dataset (head + tail, the parts that matter), bringing sequences from ~24k down to 3–5k tokens. At that point the fp32 logits take ~6 GB — they fit comfortably in 80 GB — and the liger fused loss was no longer needed. In fact it was the very cause of the last failure: disabled, standard DPO ran clean.

The final smoke test: 8 steps, loss trending down, positive rewards/margins

, accuracies

0.75–0.88. The model preferred the chosen. Green light.

The full run trained in the background on the A100 (detached, to survive SSH drops), monitored step by step:

70 steps, 1 epoch over 559 pairs, 4h11m, zero OutOfMemory from start to finish — even when VRAM read 96% (a false alarm: that was PyTorch's stable reserved pool, not a live allocation at the edge; the paged 8-bit optimizer spills to system RAM as a relief valve). Margins above 1.0 and accuracies at 88% say the preference signal was learned cleanly.

The adapter was merged, converted to GGUF q4_K_M, and served by Ollama as qwen3-coder-wuic:30b-dpo

. Then came the eval gate, with criteria fixed before looking at the results:

A/B on the same clean bench, three rounds per task:

| Criterion | Base | DPO | Gate | Verdict |

|---|---|---|---|---|
| PASS on HARD | 25/36 (69%) | 28/36 (78%) |

≥ 25/36 | pass | | Curriculum | 96/96 | 96/96 | 96/96 | pass | | Report guard | intact | intact | intact | pass | | Redirects/run | 0.67 | 0.83 | < 0.47 | fail | | Zero-redirect runs | 53% | 39% | > 53% | fail |

The capability gains are real: Q02 went from 0/3 to 3/3 — a task the base never solved — and Q05 from 1/3 to 2/3, with zero forgetting. And on Q01, the flagship churn case the pairs were built around, redirects fell from 9 to 6.

But the primary metric — the whole reason we ran DPO — moved the wrong way: +25% redirects overall. Digging in, the number is less damning than it looks. The base scores zero redirects on Q02 because it fails cleanly; the DPO model racks up four because it fights and wins (0/3 → 3/3). A redirect count that punishes engaging with hard tasks is a metric with a blind spot — a flaw we only saw because the gate forced us to stare at it.

The pre-registered rule, though, was: no automatic adoption unless the gate clears. It didn't clear. The base stays the plugin's default; the DPO model sits one setting away, and switching in either direction is trivial — exactly the cheap, reversible outcome the gate was designed to buy.

One epilogue. On a different job — single-shot tool routing for the in-product RAG chatbot, where there are no guards and no multi-turn churn to fall back on — the same DPO model scored 96/102 against the base's 88/102 at identical speed, and that gate it cleared. It now ships as the chatbot's default brain. The model didn't fail; it turned out to be better at a different job than the one we trained it for.

Because that's the real point of this whole story. Not "we fine-tuned a model," but: we built a ruler that doesn't lie, we cleaned the test bench, we fixed the criteria before looking — and we let the numbers decide. Twice.

The model is public. qwen3-coder-wuic:30b-dpo is on Hugging Face as a GGUF (q4_K_M, ~19 GB) — pull it straight into Ollama and run it on your own 24 GB GPU. The full setup, end to end — the Ollama tuning, the WUIC backend config, and the WUIC Assistant VS Code extension — is on the Run the WUIC coding model locally page. No API calls, no data leaving your machine.