effect-qwen36-35b-react-lora

Effect↔React Seam Qwen3.6-35B-A3B — React/Seam LoRA (MLX, rank-8, unfused)

A local LoRA adapter that teaches Qwen3.6-35B-A3B-4bit to write the thin React front-end + the React↔Effect seam of a modern typed-functional stack: React 19 + Mantine v9 view, TanStack Query/Start, and the Effect v4 boundary ([email protected], effect/unstable/*). Built $0-local on a single Apple M5 Max (48 GB). It is the React half of a dual-LoRA library that shares ONE base with the Effect "write" LoRA — the two are hot-swapped in-process behind one server.

Architecture it bakes in: Contract = Effect Schema · Logic = Effect · View = React (NO business logic in .tsx) · typed errors carried across the server-fn boundary as a Schema discriminated union (pattern A) or reconstructed client-side via effect/unstable/rpc (pattern B). It writes idiomatic v4 seam code: Context.Service (not Context.Tag), Result (not Either), queryOptions/mutationOptions, useMutation().isPending (v5), Schema.TaggedErrorClass for RPC contract errors — no throw across boundaries.

Honest framing. This is a rank-8 adapter (apply unfused on the base), not a fused model. The headline numbers below are the served PRODUCT (best-of-N sampling + a real tsc --strict gate with the Mantine/TanStack/Effect type roots), not bare greedy weights — tsc is the only arbiter. The win is concentrated where it matters: the seam, which the base model gets wrong.

What it is

• Base: mlx-community/Qwen3.6-35B-A3B-4bit (qwen3_5_moe, 35.9B total / ~3B active).
• Adapter (react_v1s42_i100): rank-8 LoRA (seed 42, the val-minimum checkpoint), warm-started from the shared CPT base, SFT on 94 tsc+LSP-gated pairs (38 react-view + 56 seam) covering all 11 seam topics.
• This repo: the LoRA adapter (adapter_config.json + adapters.safetensors, the 408-tensor delta). Apply it unfused with mlx_lm.

Eval — frozen 26-task React/seam held-out, served product (best-of-N + real tsc)

The held-out set (eval/tasks_heldout_react.jsonl, 11 react-view + 15 seam) is frozen; each task is routed to the gate it must pass (react-view → React+Mantine JSX gate; seam → Effect+TanStack+Mantine gate). The same prompts are used for the base and the adapter, so any delta is the adapter's baked-in idioms.

+11 compiled / 0 regressions — every gain is on the SEAM lane (base 2/15 → 13/15; react-view 11/11 for both, i.e. the base already handles plain Mantine views, and the adapter adds the React↔Effect boundary it could not write). Measured at best-of-16; the live server runs best-of-8 with the same gate.

How it's served

The production server (serve/serve.py in the training repo) routes by intent: seam / react → this adapter; Effect business-logic / edits → the Effect "write" champion — both LoRAs on one resident base, hot-swapped in-process (a load_weights weight-overwrite; the two adapters share an identical tensor namespace). Each request is best-of-N sampled and kept only if it passes the lane's real tsc gate.

Usage (mlx_lm, Apple Silicon)

pip install mlx-lm
python -c "
from mlx_lm import load, generate
# apply the adapter unfused on the base
model, tok = load('mlx-community/Qwen3.6-35B-A3B-4bit', adapter_path='./')
prompt = 'Write a TanStack queryOptions factory userQuery(id) that fetches a User via a createServerFn running an Effect, returning a Result discriminated union. Output one tsx block.'
msgs = [{'role':'user','content':prompt}]
text = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=False)
print(generate(model, tok, text, max_tokens=640, verbose=True))
"

Tip: for production, sample N times and keep the first output that passes tsc --strict against your React/Mantine/TanStack/Effect type roots — that is the served product, not bare greedy.

Limitations

• Rank-8 adapter — needs the exact base above; greedy single-shot is weaker than the best-of-N product.
• Targets [email protected] + Mantine v9 + TanStack Query v5 / Start + React 19; other versions may drift.
• View tasks must be thin — this is deliberately NOT trained to put business logic in .tsx.
• MLX format → Apple Silicon. Reasoning disabled (enable_thinking=False); it's a direct code generator.

Companion: the Effect "write" LoRA / fused champion (jrad123777/effect-qwen36-35b-write-lora).

Built $0-local. Trained and evaluated against the installed .d.ts with tsc --strict as the only arbiter, and documented honestly.