phi-3.5-mini-instruct-bf16-mlx-cba/README.md

---
library_name: mlx
license: mit
license_link: https://huggingface.co/microsoft/Phi-3.5-mini-instruct/resolve/main/LICENSE
language:
- multilingual
pipeline_tag: text-generation
base_model: microsoft/Phi-3.5-mini-instruct
tags:
- nlp
- code
- mlx
- quantization
- bias-evaluation
---

# phi-3.5-mini-instruct-bf16 (MLX, CBA artifact)

MLX-format BF16 (uncompressed baseline) variant of [`microsoft/Phi-3.5-mini-instruct`](https://huggingface.co/microsoft/Phi-3.5-mini-instruct).

This is one of the **15 model artifacts** from the paper:

> **Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels**
> Plawan Kumar Rath, Rahul Maliakkal. *IEEE Cloud Summit 2026*.
> Code: <https://github.com/plawanrath/compression-bias-amplification>
> arXiv: <https://arxiv.org/abs/2605.15208>

## Quantization

This is the **BF16 baseline** used as the uncompressed reference in the paper. Weights have been re-serialized via `mlx_lm.convert` (no quantization) so this directory is loadable directly by MLX without an extra conversion step.

## How this artifact was produced

```bash
python -m mlx_lm.convert \
    --hf-path microsoft/Phi-3.5-mini-instruct \
    --mlx-path ./phi-3.5-mini-instruct-bf16 \

```

This is the **exact** artifact used to produce the inference results in §4.3 of the paper (911,100 records over BBQ ambiguous, 5 seeds × 12,148 items × 15 configs).

## Usage (MLX)

```bash
pip install mlx-lm
```

```python
from mlx_lm import load, generate

model, tokenizer = load("plawanrath/phi-3.5-mini-instruct-bf16-mlx-cba")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Hello!"}],
    add_generation_prompt=True,
    tokenize=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=128))
```

Or via CLI:

```bash
mlx_lm.generate --model plawanrath/phi-3.5-mini-instruct-bf16-mlx-cba --prompt "Hello!"
```

## Paper findings relevant to this variant

The paper documents a **dose-response** relationship between quantization aggressiveness and emergent stereotypical behavior on BBQ ambiguous questions:

| Variant | % of BF16-unbiased items that became biased |
|---|---|
| Q8 | 0.1–0.9% |
| Q6 | 0.3–1.3% |
| Q4 | 2.2–5.6% |
| Q3 | 6.0–21.1% |

These changes are largely **invisible to perplexity** (<0.5% shift at Q8, <3% at Q4 across all three families). Treat any deployment of compressed instruction-tuned models on fairness-sensitive tasks accordingly.

## Model details

- **Base model:** [`microsoft/Phi-3.5-mini-instruct`](https://huggingface.co/microsoft/Phi-3.5-mini-instruct)
- **Family:** Phi-3
- **Parameters:** 3.8B
- **Precision:** BF16 (uncompressed baseline)
- **Format:** MLX (Apple Silicon)
- **Conversion framework:** [`mlx-lm`](https://github.com/ml-explore/mlx-lm)

## License

Inherited from the base model (`mit`). See the upstream model page for the full license text.

## Citation

```bibtex
@inproceedings{rath2026quantization,
  title     = { Quantization Undoes Alignment: Bias Emergence in Compressed LLMs Across Models and Precision Levels },
  author    = {Rath, Plawan Kumar and Maliakkal, Rahul},
  booktitle = { IEEE Cloud Summit 2026 },
  year      = {2026},
  eprint    = {2605.15208},
  archivePrefix = {arXiv},
  url       = {https://arxiv.org/abs/2605.15208}
}
```