Instructions to use jrad123777/effect-qwen36-35b-write-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use jrad123777/effect-qwen36-35b-write-lora with MLX:
# Make sure mlx-lm is installed # pip install --upgrade mlx-lm # if on a CUDA device, also pip install mlx[cuda] # Generate text with mlx-lm from mlx_lm import load, generate model, tokenizer = load("jrad123777/effect-qwen36-35b-write-lora") prompt = "Once upon a time in" text = generate(model, tokenizer, prompt=prompt, verbose=True) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- MLX LM
How to use jrad123777/effect-qwen36-35b-write-lora with MLX LM:
Generate or start a chat session
# Install MLX LM uv tool install mlx-lm # Generate some text mlx_lm.generate --model "jrad123777/effect-qwen36-35b-write-lora" --prompt "Once upon a time"
add model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
base_model: mlx-community/Qwen3.6-35B-A3B-4bit
|
| 3 |
+
library_name: mlx
|
| 4 |
+
pipeline_tag: text-generation
|
| 5 |
+
tags:
|
| 6 |
+
- lora
|
| 7 |
+
- mlx
|
| 8 |
+
- effect-ts
|
| 9 |
+
- typescript
|
| 10 |
+
- code-generation
|
| 11 |
+
license: apache-2.0
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# effect-qwen36-35b-write-lora — Effect v4 TypeScript champion LoRA (`v7s43_i200`)
|
| 15 |
+
|
| 16 |
+
Rank-8 LoRA adapter on **`mlx-community/Qwen3.6-35B-A3B-4bit`** that writes idiomatic
|
| 17 |
+
**Effect v4** (`effect@4.0.0-beta.8x`) TypeScript: `Context.Service`/`Layer` services,
|
| 18 |
+
`Effect.gen` pipelines, `Schema`, `Data.TaggedError`, and typed-error seams — verified by a
|
| 19 |
+
best-of-N + TypeScript-compiler product harness, not by eyeballing.
|
| 20 |
+
|
| 21 |
+
## Result — the served PRODUCT (not raw greedy)
|
| 22 |
+
- **best-of-16 + tsc gate: 23/24** on a frozen held-out set. This is the deliverable metric.
|
| 23 |
+
- Judged on the served product pipeline (sample N → import-resolver → `tsc --strict`), **never
|
| 24 |
+
raw single-greedy**: raw greedy is ceiling-blocked at ~9.67/24; the compiler verifier is what
|
| 25 |
+
lifts it to 23/24. (A perfect 0/1 verifier means a returned "compiles" is genuinely verified.)
|
| 26 |
+
|
| 27 |
+
## Adapter
|
| 28 |
+
- LoRA: rank 8, scale 20, dropout 0, 16 layers; lr 1e-5; warm-started from a continued-pretrain
|
| 29 |
+
adapter. Apply **unfused** on top of the 4-bit base.
|
| 30 |
+
- `adapters.safetensors` = 1,025,848,752 bytes (the byte-intact champion).
|
| 31 |
+
|
| 32 |
+
## Use (MLX)
|
| 33 |
+
```bash
|
| 34 |
+
python -m mlx_lm generate \
|
| 35 |
+
--model mlx-community/Qwen3.6-35B-A3B-4bit \
|
| 36 |
+
--adapter-path . \
|
| 37 |
+
--prompt "Write an Effect service that fetches a user by id over HTTP, with a typed NotFound error."
|
| 38 |
+
```
|
| 39 |
+
For the actual deliverable, serve it behind a best-of-N + `tsc` gate (the raw adapter alone is not
|
| 40 |
+
the product). The same champion also serves a surgical **EDIT** lane (best-of-N SEARCH/REPLACE →
|
| 41 |
+
apply → `tsc` gate), tsc-verified.
|
| 42 |
+
|
| 43 |
+
## Related
|
| 44 |
+
- Fused full model: `jrad123777/effect-qwen36-35b-mlx` · GGUF: `jrad123777/effect-qwen36-35b-gguf`
|
| 45 |
+
- Companion React+seam LoRA: `jrad123777/effect-qwen36-35b-react-lora`
|