Modelpublic

albertobarnabo/fiduciary-qwen3-4b-GGUF

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes106downloads

Model Card

fiduciary-qwen3-4b-GGUF

GGUF quants of [Fiduciary](https://huggingface.co/albertobarnabo/fiduciary-qwen3-4b) — a local-first personal-finance advisor LLM (Qwen3-4B-Instruct-2507 fine-tune), for llama.cpp, Ollama, LM Studio, Jan, and everything else in the GGUF universe. Your portfolio questions never leave your machine.

⚠️ Educational tool, not financial advice. Fiduciary explains concepts, reads the portfolio you show it, and talks through trade-offs like a patient advisor — it does not know your full situation and it can be wrong. Decisions are yours.

How do I run a local financial-advisor LLM?

Ollama (fastest path):

bash

ollama run hf.co/albertobarnabo/fiduciary-qwen3-4b-GGUF:Q4_K_M

LM Studio: search fiduciary in the model browser, pick Q4_K_M.

llama.cpp:

bash

llama-cli -hf albertobarnabo/fiduciary-qwen3-4b-GGUF:Q4_K_M \
  -p "I'm 28, saving 800€/month. Explain index funds vs picking stocks, for my situation."

Which file should I download?

file	size	use it when
`fiduciary-qwen3-4b-Q4_K_M.gguf`	2.5 GB	default — best size/quality balance, runs on 8 GB RAM
`fiduciary-qwen3-4b-Q5_K_M.gguf`	2.9 GB	a bit more quality headroom, 8–16 GB RAM
`fiduciary-qwen3-4b-Q8_0.gguf`	4.3 GB	near-lossless, 16 GB+ RAM

Need F16 or another quant for re-quantizing? Open a discussion — happy to add it.

What is this model?

A LoRA fine-tune (rank 16, 16 layers, prompt-masked so only advisor turns train) of Qwen3-4B-Instruct-2507, trained with mlx-lm on synthetic advisor dialogues: portfolio review conversations, tool-calling traces (live prices/news), and financial-literacy explanations. Final validation loss 0.046 — a training-sanity signal on a small validation set, not a quality benchmark; no standardized finance benchmark is claimed here.

These GGUFs were made the standard QLoRA way: adapters (trained against the 4-bit MLX base) fused onto the full-precision Qwen3-4B-Instruct-2507, then quantized with llama.cpp b10092. The chat template is embedded.

The full agent runtime — local portfolio.json, price/news tools, the loop that makes it an advisor rather than a chatbot — lives in the GitHub repo. The weights are the brain; the repo is the owl's perch. 🦉

The Fiduciary family

artifact	for
fiduciary-qwen3-4b	Apple silicon / MLX (4-bit, 2.3 GB)
this repo	everyone else: Ollama, LM Studio, llama.cpp, Jan
fiduciary-qwen3-4b-lora	the raw LoRA adapter (56 MB) — stack it on your own Qwen3-4B

Limitations

—English-first; finance terminology skews US/EU retail investing.
—A 4B model: confidently wrong sometimes — verify numbers before acting.
—Knowledge is frozen at base-model training; live data comes from the agent tools, not the weights.
—Trained on synthetic dialogues (generated + curated), not real client conversations.