Llama-3.1-8B Financial Sentiment — QLoRA

LoRA adapters fine-tuned on FinGPT/fingpt-sentiment-train with QLoRA: 4-bit NF4 frozen base + bf16 LoRA adapters (r=16, alpha=32, dropout=0.05).

Task: 3-class financial sentiment — negative / neutral / positive

Evaluation Results

Dataset	Base Accuracy	FT Accuracy	Base Macro-F1	FT Macro-F1
FPB in-domain	0.8908	0.9748	0.8765	0.9725
FiQA-SA OOD	0.8120	0.9402	0.6705	0.8335

Baseline = zero-shot Meta-Llama-3.1-8B-Instruct with the same prompt template.

Quick Start

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Meta-Llama-3.1-8B-Instruct",
    load_in_4bit=True,
    device_map="auto",
)
model     = PeftModel.from_pretrained(base, "jhon53/Llama3_1_8B_Finance_QLoRA")
tokenizer = AutoTokenizer.from_pretrained("jhon53/Llama3_1_8B_Finance_QLoRA")

Training Details

Param	Value
Base model	meta-llama/Meta-Llama-3.1-8B-Instruct
Method	QLoRA (4-bit NF4 + LoRA bf16)
LoRA rank (r)	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	q, k, v, o, gate, up, down projections
Training data	FinGPT/fingpt-sentiment-train (~76k)
Optimizer	AdamW 8-bit
Learning rate	2e-4 (cosine schedule)
Epochs	3 (early stopping, patience=3)
Loss	Completion-only cross-entropy