---
license: apache-2.0
base_model: swiss-ai/Apertus-70B-Instruct-2509
library_name: peft
tags:
- mlx
- lora
- peft
- ailiance
- apertus
- math
language:
- en
- fr
pipeline_tag: text-generation
---

# Ailiance — Apertus-70B-Instruct math LoRA

LoRA adapter fine-tuned on `swiss-ai/Apertus-70B-Instruct-2509` for **math** tasks.

> Maintained by **Ailiance** — French AI org publishing EU AI Act aligned LoRA adapters and datasets.

## Quick start (MLX)

```python
from mlx_lm import load, generate

model, tokenizer = load(
    "swiss-ai/Apertus-70B-Instruct-2509",
    adapter_path="Ailiance-fr/apertus-math-lora",
)

print(generate(model, tokenizer, prompt="..."))
```

## Training

| Hyperparameter   | Value                  |
|------------------|------------------------|
| Base model       | `swiss-ai/Apertus-70B-Instruct-2509`     |
| Method           | LoRA via `mlx-lm`      |
| Rank             | 16            |
| Scale            | 2.0           |
| Alpha            | 32           |
| Max seq length   | 1024  |
| Iterations       | 500           |
| Optimizer        | Adam, LR 1e-5          |
| Hardware         | Apple M3 Ultra 512 GB  |

## Training data lineage

Derived from the internal **eu-kiki / mascarade** curation. All upstream samples
are synthetic, permissively-licensed, or generated from Apache-2.0 base resources.
See the [Ailiance-fr catalog](https://huggingface.co/Ailiance-fr) for related cards.

## Benchmark roadmap

This LoRA has **not yet been evaluated** through `electron-bench` (the current
pipeline supports `gemma-4-E4B` base only). Training was completed with the
standard `mlx-lm` LoRA trainer (rank 16, alpha 32, scale 2.0, AdamW
LR 1e-5, 500 iters) — full hyperparameters are in the `Training` table above.

Planned evaluations:

- Perplexity on the validation split of the training data
- Functional benchmark on **apertus**-specific tasks
- Comparison vs base `swiss-ai/Apertus-70B-Instruct-2509`

Track progress: [ailiance-bench issues](https://github.com/ailiance/ailiance-bench/issues).

For reference benchmarks on the `gemma-4-E4B` base, see the
[base-vs-LoRA matrix](https://github.com/ailiance/ailiance-bench/blob/main/bench-results/compare_base_vs_lora.md).

## License chain

| Component                         | License           |
|-----------------------------------|-------------------|
| Base model (`swiss-ai/Apertus-70B-Instruct-2509`)        | apache-2.0    |
| Training data (internal Ailiance curation (synthetic + permissive sources))         | apache-2.0      |
| **LoRA adapter (this repo)**      | **apache-2.0**|

_All upstream components are Apache 2.0 / MIT — LoRA inherits permissive terms._

## EU AI Act compliance

- **Article 53(1)(c)**: training data licenses preserved (per-dataset cards declare upstream licenses).
- **Article 53(1)(d)**: training data summary — see upstream dataset cards on Ailiance-fr.
- **GPAI Code of Practice (July 2025)**: base `swiss-ai/Apertus-70B-Instruct-2509` released under apache-2.0.
- **No web scraping by Ailiance**, **no licensed data**, **no PII**.
- Upstream Stack Exchange content (where applicable) is CC-BY-SA-4.0 and propagates to this adapter.

## License

LoRA weights: **apache-2.0** — see License chain table above for derivation rationale.

## Citation

```bibtex
@misc{ailiance_apertus_math_2026,
  author    = {Ailiance},
  title     = {Ailiance — Apertus-70B-Instruct math LoRA},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/Ailiance-fr/apertus-math-lora}
}
```

## Related

See the full [Ailiance-fr LoRA collection](https://huggingface.co/Ailiance-fr).


## Bench comparison (2026-05-11)

### Base model (Apertus-70B-Instruct-2509) capability

| Task | Score | Notes |
|---|---:|---|
| ARC-Easy acc / acc_norm | **0.81 / 0.77** | W3 lm-eval-harness BF16 |
| GSM8K-CoT | TIMEOUT (1800s budget) | base 70B BF16 too slow for CoT |
| MMLU-Pro Computer Science | TIMEOUT | |

### This LoRA (tuned) — bench PENDING

Production usage: served via gateway alias `ailiance-apertus-<domain>` on
<https://www.ailiance.fr> through the Apertus multi-LoRA hot-swap server
(Studio :9322, 1 base + 10 LoRA dynamic swap, ~40GB VRAM).

## Upstream base model — official evaluations

This LoRA fine-tunes [`swiss-ai/Apertus-70B-Instruct-2509`](https://huggingface.co/swiss-ai/Apertus-70B-Instruct-2509),
the EU-sovereign open-source LLM released by the Swiss AI Initiative. Below are
the **official scores** reported in the [Apertus Tech Report](https://arxiv.org/abs/2509.14233)
on a suite of multilingual reasoning benchmarks.

| Model                       | Avg   | ARC   | HellaSwag | WinoGrande | XNLI  | XCOPA | PIQA  |
|-----------------------------|------:|------:|----------:|-----------:|------:|------:|------:|
| **Apertus-70B** (this base) | 67.5  | 70.6  | 64.0      | 73.3       | 45.3  | 69.8  | 81.9  |
| Apertus-8B                  | 65.8  | 72.7  | 59.8      | 70.6       | 45.2  | 66.5  | 79.8  |
| Llama3.1-70B                | 67.3  | 74.4  | 56.5      | 79.4       | 44.3  | 66.7  | 82.3  |
| Qwen2.5-72B                 | 69.8  | 76.2  | 67.5      | 78.0       | 46.9  | 68.2  | 82.0  |
| OLMo2-32B                   | 67.7  | 76.2  | 66.7      | 78.6       | 42.9  | 60.1  | 82.1  |
| EuroLLM-9B                  | 62.8  | 67.9  | 57.9      | 68.8       | 41.5  | 61.1  | 79.6  |

Many additional benchmark evaluations (pretraining/post-training phases,
multilingual in ~100 languages, long-context) are in Section 5 of the
[Apertus Tech Report](https://arxiv.org/abs/2509.14233).

**Source:** [official Apertus-70B-Instruct-2509 model card](https://huggingface.co/swiss-ai/Apertus-70B-Instruct-2509).

> **Reading these alongside this LoRA:** Apertus-70B is EU AI Act-compliant
> (`Apertus_EU_Code_of_Practice.pdf`, `Apertus_EU_Public_Summary.pdf` included
> in upstream weights). This LoRA inherits that compliance plus the
> general-capability floor shown above, then adds domain specialization.