---
library_name: transformers
pipeline_tag: text-generation
tags:
- arithmetic
- interpretability
- arxiv:2405.14813
---

# Fixed-width addition transformer

Run `s85nnxtf` is a 1-block, bias-free causal transformer trained for
4-digit base-10 addition. Operands are zero-padded and answers use
5 digits, retaining overflow.

## Results

| Metric | Value |
| --- | ---: |
| Validation loss | 0.003520 |
| Validation generated-token accuracy | 99.85% |
| Validation exact-answer accuracy | 99.32% |
| No-carry exact-answer accuracy | 97.27% |
| Single-carry exact-answer accuracy | 100.00% |
| Multiple-carry exact-answer accuracy | 98.44% |
| Carry-chain exact-answer accuracy | 97.27% |

## Training configuration

- Updates: 10000
- Optimizer: muon
- Muon peak learning rate: 0.02
- AdamW peak learning rate: 0.0003
- Weight decay: 0.01
- Warmup updates: 100
- Minimum learning-rate ratio: 0.1
- Initialization: normal
- Seed: 0
- Source commit: `unavailable`

The complete resolved configuration, environment, metrics, source snapshot, and checkpoints are
available in [`training/`](./training/). Machine-readable hashes and metrics are in
[`export_manifest.json`](./export_manifest.json).

## Loading

This repository contains custom Transformers code. For reproducible or security-sensitive use,
pin the commit revision printed by the uploader.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

revision = "PINNED_COMMIT_HASH"
tokenizer = AutoTokenizer.from_pretrained(
    "OWNER/REPO", trust_remote_code=True, revision=revision
)
model = AutoModelForCausalLM.from_pretrained(
    "OWNER/REPO", trust_remote_code=True, revision=revision
)
inputs = tokenizer("0000 + 0000 =", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=model.config.answer_digits, do_sample=False)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```

## Intended use and limitations

This model is intended for mechanistic-interpretability research on its configured fixed-width
addition task. It is not a general arithmetic system: inputs outside the configured grammar or
width are unsupported, and generated answers must not be treated as reliable calculations.