README.md

---
pretty_name: "PixCell"
license: mit
base_model: Qwen/Qwen3.6-35B-A3B
library_name: peft
pipeline_tag: image-text-to-text
language:
  - en
datasets:
  - qpaig-mit/pixcell
tags:
  - lora
  - grpo
  - reinforcement-learning
  - image-to-code
  - code-generation
  - photonics
  - gds
  - gdsfactory
---

# PixCell — blind photonic reconstructor

A LoRA adapter (r=32, all-linear) for `Qwen/Qwen3.6-35B-A3B` that converts a
black-and-white raster of a photonic component plus its physical footprint
into a primitive-only [GDSFactory](https://github.com/gdsfactory/gdsfactory)
Python program. It was trained with GRPO alone — no supervised stage — on the
[PixCell curriculum](https://huggingface.co/datasets/qpaig-mit/pixcell)
(levels L0→L4) against a deterministic visual verifier that renders each
program and scores it at calibrated absolute scale. Companion model release
of *From Pixels to PCells* ([QPG-MIT/PixCell](https://github.com/QPG-MIT/PixCell)).

## This model is an iterative system

The model is one half of an attempt–measure–revise loop, and the loop is how
you run it: sample 8 attempts, render and score each with the deterministic
verifier, keep the champion, then revise the champion for 3 rounds of 4
revisions under number-free feedback — a closeness bucket plus whether the
rendering has EXTRA or MISSING material, read from boundary-chamfer
asymmetry. The verifier only measures; the model does all the searching. The
loop and verifier ship in the repo
([`rl/track_a/agent_loop.py`](https://github.com/QPG-MIT/PixCell/blob/main/rl/track_a/agent_loop.py),
[`rl/common/evaluator.py`](https://github.com/QPG-MIT/PixCell/blob/main/rl/common/evaluator.py)).

On the repo's 8-device held-out benchmark (`data/benchmark/final_1..8` —
real devices from published papers, never seen in training), the loop
reaches **mean champion IoU 0.491**:

| final_1 | final_2 | final_3 | final_4 | final_5 | final_6 | final_7 | final_8 |
|---|---|---|---|---|---|---|---|
| 0.518 | 0.507 | 0.336 | 0.651 | 0.293 | 0.716 | 0.489 | 0.418 |

Used instead as a bare one-shot generator (single pass, T=1.0), expect mean
IoU 0.23 with 39/64 attempts executable — the base model scores 0/64.

## Prompt contract

The deployment prompt is the training prompt, unchanged
([`rl/common/prompt.py`](https://github.com/QPG-MIT/PixCell/blob/main/rl/common/prompt.py),
contract `pixcell-direct-reconstruction-v3`): the device raster (longest side
≤1440 px) and its footprint in μm, non-thinking chat template, ≤4096 new
tokens, temperature 1.0. The model answers with one Python program over the
seven-primitive GDSFactory catalog that writes `device.gds`.

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-35B-A3B", torch_dtype="bfloat16")
model = PeftModel.from_pretrained(base, "qpaig-mit/pixcell")
# ...then drive it with the repo's agent loop; a lone forward pass is not the system.
```

## Training

GRPO (importance-sampling loss, KL 0, group size 8, 8 groups/step, lr 1e-5,
temperature 1.0) over the curriculum L0→L4, 30 steps per level with 80/20
replay of earlier levels; reward `shaped_v3b` — rectangle-baseline-normalized
IoU and Dice plus a boundary-chamfer teacher over a 0.05 validity floor,
recomputed from rendered artifacts by the deterministic evaluator
([`rl/track_a/reward.py`](https://github.com/QPG-MIT/PixCell/blob/main/rl/track_a/reward.py)).
Trained with [Tinker](https://thinkingmachines.ai/tinker). The full evidence
record (benchmark JSONs, loop trajectories, champion programs, gallery) is
versioned at
[`data/training/qwen-rl-v2/`](https://github.com/QPG-MIT/PixCell/tree/main/data/training/qwen-rl-v2).