Paper: arXiv 2601.10043

Authors: Zhiming Lian

Abstract

Particularly, financial named-entity recognition (NER) is one of the many important approaches to translate unformatted reports and news into structured knowledge graphs. However, free, easy-to-use large language models (LLMs) often fail to differentiate organisations as people, or disregard an actual monetary amount entirely. This paper takes Meta’s Llama 3 8B and applies it to financial NER by combining instruction fine-tuning and Low-Rank Adaptation (LoRA). Each annotated sentence is converted into an instruction-input-output triple, enabling the model to learn task descriptions while fine-tuning with small low-rank matrices instead of updating all weights. Using a corpus of 1,693 sentences, our method obtains a micro-F1 score of 0.894 compared with Qwen3-8B, Baichuan2-7B, T5, and BERT-Base. We present dataset statistics, describe training hyperparameters, and perform visualizations of entity density, learning curves, and evaluation metrics. Our results show that instruction tuning combined with parameter-efficient fine-tuning enables state-of-the-art performance on domain-sensitive NER.

Complexity vs Empirical Score

  • Math Complexity: 3.0/10
  • Empirical Rigor: 7.5/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper’s mathematical content is limited to describing standard neural architectures and LoRA decomposition, lacking advanced derivations or novel theoretical frameworks. Empirical rigor is high, featuring a concrete dataset with 1,693 annotated sentences, comparative benchmarks against multiple models, hyperparameter reporting, and visualizations of learning curves and evaluation metrics.

Research Flowchart

  flowchart TD
  A["Research Goal: Financial Named-Entity Recognition<br>using LLaMA-3-8B & LoRA"] --> B["Data Preparation<br>1,693 annotated sentences<br>Instruction-Input-Output format"]
  B --> C["Methodology<br>Instruction Fine-tuning + LoRA"]
  C --> D["Training Process<br>Parameter-efficient updates<br>Low-rank matrices"]
  D --> E["Computational Evaluation<br>Micro-F1 Score = 0.894<br>vs. Qwen3-8B, Baichuan2-7B, BERT"]
  E --> F["Key Outcomes<br>State-of-the-art Financial NER<br>Efficient domain adaptation"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes an instruction‑tuned LoRA adaptation of Llama 3 for financial named‑entity recognition, fine‑tuning only low‑rank matrices while keeping the base model frozen.

Source sentence
We presented an instruction-tuned LoRA adaptation of Llama 3 for financial named-entity recognition. By transforming annotated sentences into instruction–input–output triples and fine-tuning only low-rank matrices, the model achieves a micro-F1 of 0.894

Assumes 2

  • The frozen pre‑trained Llama 3 model provides sufficient linguistic knowledge, so only low‑rank adapters need to be trained. model
    Source sentence
    We fine‑tune only the low‑rank matrices inserted into each attention projection while keeping the base model frozen.
  • A corpus of 1,693 annotated sentences is representative enough to evaluate financial NER performance. data
    Source sentence
    Using a corpus of 1 693 sentences our method can obtain a micro‑F1 of 0.894 in comparison to Qwen3-8B, Baichuan2-7B, T5 and BERT-Base.

Data

  • Financial NER Dataset text-news unknown
    Source sentence
    Table 2 shows the financial dataset statistic, which shows the seven entities. The company entity has the most samples around 1033, while the product entity has the fewest samples around 226.

Costs

Not applicable The paper reports no trading or execution result.

Finds 1

  • The LoRA‑based fine‑tuning accelerates convergence, showing lower training loss in few iterations compared to full‑parameter fine‑tuning. performance
    Source sentence
    The figure 4 shows that the LoRA finetuning can accelerate the model convergence in a few iterations, which show the superiority of instruction finetuning.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 10 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (9)

  1. FinLoRA: Benchmarking LoRA Methods for Fine-Tuning LLMs on Financial Datasets (2025) — link
  2. Financial Named Entity Recognition: How Far Can LLM Go? (2025) — link
  3. EGGesture: Entropy-Guided Vector Quantized Variational AutoEncoder for Co-Speech Gesture Generation (2024) — link
  4. Retrieval Augmented Instruction Tuning for Open NER with Large Language Models (2024) — link
  5. Baichuan2-Sum: Instruction Finetune Baichuan2-7B Model for Dialogue Summarization (2024) — link
  6. LoRA: Low-Rank Adaptation of Large Language Models (2021) — link
  7. Attention is All you Need (2017) — link
  8. “FODA-PGforEnhancedMedicalImagingNar-rativeGeneration:AdaptiveDifferentiationofNormalandAbnormalAttributes,”
and 1 more
  1. “AdvancingEntityRecognitioninBiomedicinevia Instruction Tuning of Large Language Models,”

0 of 9 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.