README.md

---
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
library_name: transformers
pipeline_tag: text-generation
language:
- en
tags:
- reversible-circuit
- quantum-circuit-synthesis
- code
- reasoning
- ecdsa-fail
- verifier-guided
---

# reversible-circuit-coder-1.5b

**A 1.5B model fine-tuned to synthesize cheap, correct reversible quantum circuits — and an honest case
study in where small-model imitation/RL/reasoning hits a wall on algorithmic tasks.**

This model designs reversible (quantum) circuits for the [ECDSA.fail](https://ecdsa.fail) secp256k1
point-addition challenge and the broader task of **verifier-guided, cost-minimizing reversible-circuit
optimization**: given a target reversible function, emit a circuit that is correct on every input,
reversible, phase-clean, and ancilla-clean, at the lowest cost (Toffoli count × peak qubit width).

- **Developed by:** Dennison Bertram (built autonomously with Claude Code)
- **Base model:** `Qwen/Qwen2.5-Coder-1.5B-Instruct` (Apache-2.0)
- **License:** Apache-2.0
- **Repository (full pipeline, verifier, data factory, eval, honest writeup):**
  [github.com/dennisonbertram/reversible-circuit-llm](https://github.com/dennisonbertram/reversible-circuit-llm)

## Model description
The training signal comes from a **microsecond-exact verifier** (bit-identical to the challenge's Rust
simulator). Rather than fine-tuning on textbook examples, a **verifier-gated search engine** produces
*near-optimal* circuits (~0.54× the Toffoli cost of textbook references), and the model is SFT'd
(LoRA) on **24,545** such optimal targets across a 7-family curriculum. The model emits an op-stream in
the harness DSL: `X qT`, `CX qC qT`, `CCX qC1 qC2 qT` (Toffoli — the cost lever), `SWAP qA qB`.

## Intended uses & limitations
**Intended:** a proof-of-concept / research artifact for verifier-grounded circuit synthesis; a
generator of small reversible arithmetic/boolean circuits (use best-of-N with the open-source verifier
as an inference oracle); a teaching example for neuro-symbolic / tool-use research.

**Not intended:** a production solver. It reliably solves only the easiest tasks.

## Evaluation (honest)
Held-out reversible-circuit **synthesis**, `valid_rate` = fraction solved with best-of-16:

| model | held-out valid_rate |
|---|---|
| base Qwen2.5-Coder-1.5B | 0% (emits Python, not circuits) |
| **this model (optimal-target SFT)** | **4.8%** (solves the easiest band) |

**Key research finding:** a 7B trained identically, plus reinforcement learning (GRPO) and
reasoning chain-of-thought, **all plateau at the same ~4%**. The bottleneck is **not** data, capacity,
RL, or reasoning — it is the small model's inability to reliably *execute* multi-step symbolic
procedures (Gaussian elimination, ripple-carry) for unseen instances. It can *narrate* the algorithm
but makes *execution* errors. Even a state-externalizing tool (single gate at a time) didn't break this
zero-shot — the remaining gap is **sequential planning**. The honest next directions are tool-use with
training, frontier-scale reasoning models, and neuro-symbolic methods.

## How to use
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("dennisonb/reversible-circuit-coder-1.5b")
model = AutoModelForCausalLM.from_pretrained("dennisonb/reversible-circuit-coder-1.5b")
```
Use the system prompt + task format from the repo (`proxy/system_prompt.txt`, `proxy/sample_task.txt`),
sample best-of-N, and verify each candidate with the open-source proxy verifier (`proxy/proxy_env.py`).

## Training data
24,545 near-optimal circuit targets generated by the verifier-gated search engine over a procedurally
generated curriculum (modular adders/multipliers/inverse, controlled add/sub, GF(2) linear maps,
S-boxes; widths 2–7). Move/reasoning corpora mined from 275 accepted ECDSA.fail submissions are also in
the repo. Datasets are regenerable via the repo's scripts.

🤖 Built autonomously with Claude Code.