qwen25-14b-fin-lora — Equity-Analyst QLoRA Adapter

QLoRA adapter for Qwen/Qwen2.5-14B-Instruct, fine-tuned to answer deep financial / equity-analysis questions in a senior-analyst voice. The merged, ready-to-use model is at s-ttp/marketguru-14b — see that card for the full benchmark and details.

• Adapter: 4-bit nf4, r=32, α=64, all linear layers, 2 epochs (loss 1.45 → 0.43, token-acc ~87%).
• Data: 15,000 grounded instruction examples (fact-sheet → ~450–550-word analyst note) distilled via Gemini 2.5 Flash from a point-in-time, survivorship-free market panel.
• Result: +1.19 overall lift vs base (neutral Kimi K2.6 judge); near-frontier equity analysis at ~1% of frontier cost.

Not for return prediction or investment advice; can hallucinate — verify numbers.

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-14B-Instruct", device_map="auto", load_in_4bit=True)
model = PeftModel.from_pretrained(base, "s-ttp/qwen25-14b-fin-lora")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-14B-Instruct")