README.md

---
license: apache-2.0
language:
  - en
base_model: Qwen/Qwen3-4B-Instruct-2507
pipeline_tag: text-generation
library_name: mlx
tags:
  - mlx
  - apple-silicon
  - qwen3
  - finance
  - financial-advisor
  - financial-literacy
  - personal-finance
  - local-llm
  - privacy
  - self-hosted
  - lora
---

> **Nota sui parametri**: questo repo contiene i pesi **Qwen3-4B quantizzati 4-bit via MLX** (uint32 packing). Il contatore di parametri di Hugging Face legge i tensori impacchettati e mostra un valore errato: il modello ha **4B parametri**, non 0.6B.

# fiduciary-qwen3-4b 🦉

**A personal-finance advisor LLM that lives on your Mac — not in someone else's cloud.**
This is the **MLX 4-bit** build (2.3 GB) of Fiduciary, a Qwen3-4B-Instruct-2507
fine-tune for Apple silicon. It reads the portfolio *you* keep in a local JSON file,
explains investing concepts like a patient senior advisor, and never sends a byte
of your finances anywhere.

> ⚠️ **Educational tool, not financial advice.** Fiduciary explains and discusses;
> it does not know your full situation and can be wrong. Decisions are yours.

**Not on a Mac?** Use the [GGUF build](https://huggingface.co/albertobarnabo/fiduciary-qwen3-4b-GGUF)
for Ollama / LM Studio / llama.cpp. This repo's weights are MLX-quantized and load
with `mlx-lm` — they will **not** load with plain 🤗 `transformers`.

## Why would I run a financial advisor locally?

Because the alternative is uploading your portfolio, your balances, and your money
anxieties to an API. A 4B model on your own machine is private by construction:
you own the weights (Apache-2.0), you own the data (a plain `portfolio.json` on
disk), and you can read every tool it calls in the
[GitHub repo](https://github.com/albertobarnabo/fiduciary) — the agent runtime
that fetches live prices and news lives there, in auditable Python, not in the weights.

## How do I run it on my Mac?

```bash
pip install mlx-lm
mlx_lm.chat --model albertobarnabo/fiduciary-qwen3-4b
```

```python
from mlx_lm import load, generate
model, tokenizer = load("albertobarnabo/fiduciary-qwen3-4b")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "I'm 28 and can save 800€/month. Walk me through index funds vs stock picking for someone like me."}],
    add_generation_prompt=True, tokenize=False)
print(generate(model, tokenizer, prompt, max_tokens=512))
```

For the full advisor experience (portfolio file, live prices, news) clone the
[runtime repo](https://github.com/albertobarnabo/fiduciary) and follow its README.

## How was it trained?

LoRA (rank 16, first 16 layers) on **Qwen3-4B-Instruct-2507** with `mlx-lm` on a
16 GB MacBook — the whole fine-tune ran on consumer Apple silicon. Data: synthetic
advisor dialogues (portfolio reviews, tool-calling traces, financial-literacy
explanations), generated and curated in the runtime repo; `mask_prompt: true` so
only advisor turns contribute to the loss, expanded per assistant turn. Final
validation loss 0.046 — read that as a training-sanity signal on a small
validation set, **not** a benchmark; this model makes no benchmark claims.

## The Fiduciary family

| artifact | for |
|---|---|
| **this repo** | Apple silicon via MLX (4-bit, 2.3 GB) |
| [fiduciary-qwen3-4b-GGUF](https://huggingface.co/albertobarnabo/fiduciary-qwen3-4b-GGUF) | Ollama, LM Studio, llama.cpp, Jan — every other machine |
| [fiduciary-qwen3-4b-lora](https://huggingface.co/albertobarnabo/fiduciary-qwen3-4b-lora) | the raw LoRA adapter (56 MB) |

## Limitations

- English-first; terminology skews US/EU retail investing.
- 4B parameters: articulate, occasionally confidently wrong — verify numbers.
- Weights know nothing after the base model's cutoff; live market data comes from
  the runtime's tools, not the model.
- Trained on synthetic dialogues, not real client conversations — and it shows in
  breadth: niche instruments and tax specifics are weak spots.