jrad123777's picture
add model card
a916102 verified
|
Raw
History Blame Contribute Delete
1.92 kB
---
base_model: mlx-community/Qwen3.6-35B-A3B-4bit
library_name: mlx
pipeline_tag: text-generation
tags:
- lora
- mlx
- effect-ts
- typescript
- code-generation
license: apache-2.0
---
# effect-qwen36-35b-write-lora β€” Effect v4 TypeScript champion LoRA (`v7s43_i200`)
Rank-8 LoRA adapter on **`mlx-community/Qwen3.6-35B-A3B-4bit`** that writes idiomatic
**Effect v4** (`effect@4.0.0-beta.8x`) TypeScript: `Context.Service`/`Layer` services,
`Effect.gen` pipelines, `Schema`, `Data.TaggedError`, and typed-error seams β€” verified by a
best-of-N + TypeScript-compiler product harness, not by eyeballing.
## Result β€” the served PRODUCT (not raw greedy)
- **best-of-16 + tsc gate: 23/24** on a frozen held-out set. This is the deliverable metric.
- Judged on the served product pipeline (sample N β†’ import-resolver β†’ `tsc --strict`), **never
raw single-greedy**: raw greedy is ceiling-blocked at ~9.67/24; the compiler verifier is what
lifts it to 23/24. (A perfect 0/1 verifier means a returned "compiles" is genuinely verified.)
## Adapter
- LoRA: rank 8, scale 20, dropout 0, 16 layers; lr 1e-5; warm-started from a continued-pretrain
adapter. Apply **unfused** on top of the 4-bit base.
- `adapters.safetensors` = 1,025,848,752 bytes (the byte-intact champion).
## Use (MLX)
```bash
python -m mlx_lm generate \
--model mlx-community/Qwen3.6-35B-A3B-4bit \
--adapter-path . \
--prompt "Write an Effect service that fetches a user by id over HTTP, with a typed NotFound error."
```
For the actual deliverable, serve it behind a best-of-N + `tsc` gate (the raw adapter alone is not
the product). The same champion also serves a surgical **EDIT** lane (best-of-N SEARCH/REPLACE β†’
apply β†’ `tsc` gate), tsc-verified.
## Related
- Fused full model: `jrad123777/effect-qwen36-35b-mlx` Β· GGUF: `jrad123777/effect-qwen36-35b-gguf`
- Companion React+seam LoRA: `jrad123777/effect-qwen36-35b-react-lora`