GGUF (llama.cpp) build of vinmlops/finbot-qwen3-1.7b-baseline,
quantized to q8_0 for efficient CPU inference. This is the artifact served in-cluster
(via llama.cpp) in the MLOps pipeline (data curation → fine-tuning → evaluation gate →
in-cluster serving → monitoring).
Base of this build: vinmlops/finbot-qwen3-1.7b-baseline (QLoRA fine-tune of Qwen/Qwen3-1.7B, merged)
Format: GGUF, quantization q8_0 (~1.8 GB; high quality, CPU-friendly)
Intended use
Educational explanations of general finance concepts (IRAs, ETFs, compound interest,
diversification, etc.). Not a source of personalized financial advice.
Honest by design
Trained to decline predicting prices, guaranteeing returns, or giving personalized advice.
Not a licensed financial advisor.
Status & known limitations (v1 — baseline)
This is a v1 baseline release. Current quality is average / acceptable for an
educational demo, but not production-grade:
Answers are generally coherent and the honesty guardrails work well, but the model
can be verbose and is not consistently accurate on finer factual details.
These characteristics trace primarily to the training data: finance-alpaca is
largely sourced from public finance forums, whose style is conversational and whose
factual quality is uneven. As expected in supervised fine-tuning, the model reflects
the distribution of its training data — model quality is bounded by data quality.
Planned improvements (future v2)
A future revision will focus specifically on data quality, the highest-impact lever:
Stricter answer-quality curation (removing rambling, first-person, and low-signal responses;
filtering answers with unverifiable claims or links).
A smaller, cleaner, higher-signal training set (quality over quantity).
Re-tuning of training and decoding settings based on evaluation results.
The v1 release deliberately prioritizes a complete, working, observable pipeline over
peak model accuracy; model-quality improvements are scheduled as a focused follow-up.
