fotapol

qwen3-1.7b-financial-qa-lora

Deploy Dedicated

README

License: apache-2.0

Model details

Base model: Qwen/Qwen3-1.7B
Base model revision: 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e
Adapter type: LoRA
Training method: QLoRA with a 4-bit NF4 base model
Language: English
Domain: Financial question answering
Task: Context-grounded numerical question answering

Intended use

The adapter is intended for:

research on financial question answering;
experiments with numerical reasoning over financial context;
comparison between base and parameter-efficiently fine-tuned models;
integration into retrieval-augmented generation prototypes.

The model is experimental and must not be treated as financial, investment, accounting, tax, or legal advice.

Training data

The adapter was trained on normalized examples from:

FinQA;
DocFinQA.

Each training example used:

a system instruction;
gold supporting context;
a financial question;
the reference final answer.

The training data itself is not included in this repository.

Dataset licenses

FinQA: CC BY 4.0
DocFinQA: MIT

Training configuration

Table with columns: Parameter, Value
Parameter	Value
Training examples	11,713
Validation examples	1,642
Epochs	1
Maximum sequence length	4,096
Per-device batch size	1
Gradient accumulation	8
Effective batch size	8
Learning rate	2e-4
LoRA rank	16

Seven train examples, one validation example, and one final-holdout example were excluded because their rendered sequences exceeded 4,096 tokens (nine excluded examples in total).

Evaluation

Two evaluation sets were used:

Development set: 100 examples used for comparison with smaller fine-tuning experiments.
Final holdout: 1,937 eligible test examples not used for selecting the training subset size or hyperparameters.

Final holdout results

Table with columns: Metric, Base model, Fine-tuned adapter
Metric	Base model	Fine-tuned adapter
Exact match	0.15%	19.82%
Normalized exact match	0.46%	19.88%
Numeric accuracy	1.00%	20.45%

Numeric evaluation examples:

Base model: 19 correct out of 1,902
Fine-tuned adapter: 389 correct out of 1,902
Improved examples: 372
Regressed examples: 2

Numeric accuracy by dataset

Table with columns: Dataset, Base model, Fine-tuned adapter
Dataset	Base model	Fine-tuned adapter
FinQA	1.30%	25.07%
DocFinQA	0.61%	14.42%

Detailed aggregate results are available in the results/ directory.

Evaluation limitations

The reported numeric accuracy uses a deterministic custom evaluator that extracts and compares the first numeric value in each answer.

It is not the official FinQA program-execution or program-accuracy metric.

The evaluator does not fully normalize:

percentages versus decimal ratios;
thousands versus millions;
currencies and units;
multiple numeric values in one answer;
semantically equivalent answers with different scales.

The results should therefore be reported as:

Numeric accuracy using the project evaluator v1.

Usage

Install the required libraries:

bash
pip install \
  "transformers>=4.57.6" \
  "peft>=0.20.0" \
  "accelerate>=1.14.0" \
  "bitsandbytes>=0.50.0"

Load the base model and adapter:

python
import torch
from peft import PeftModel
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
)

BASE_MODEL_ID = "Qwen/Qwen3-1.7B"
BASE_MODEL_REVISION = (
    "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e"
)
ADAPTER_ID = "fotapol/qwen3-1.7b-financial-qa-lora"

tokenizer = AutoTokenizer.from_pretrained(
    BASE_MODEL_ID,
    revision=BASE_MODEL_REVISION,
)

quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
    bnb_4bit_use_double_quant=True,
)

base_model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL_ID,
    revision=BASE_MODEL_REVISION,
    quantization_config=quantization_config,
    device_map="auto",
)

model = PeftModel.from_pretrained(
    base_model,
    ADAPTER_ID,
)

model.eval()

Generate an answer:

python
context = """
Revenue increased from $12.4 million in 2024 to
$14.1 million in 2025.
"""

question = """
What was the percentage increase in revenue?
"""

messages = [
    {
        "role": "system",
        "content": (
            "Answer the financial question using only the "
            "provided context. Return only the final answer."
        ),
    },
    {
        "role": "user",
        "content": (
            f"Context:\n{context}\n\n"
            f"Question:\n{question}"
        ),
    },
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)

inputs = tokenizer(
    prompt,
    return_tensors="pt",
).to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=64,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

answer = tokenizer.decode(
    output[0, inputs["input_ids"].shape[1]:],
    skip_special_tokens=True,
)

print(answer.strip())

Limitations

The adapter was trained only on English data.
It was trained with gold supporting context, not with retrieved context.
It does not perform document retrieval.
Numerical calculations can still be incorrect.
DocFinQA performance remains lower than FinQA performance.
The model may return plausible but unsupported financial statements.
The training context was limited to 4,096 tokens.
Evaluation used one custom metric implementation.
The adapter has not been evaluated for financial advice or production decision-making.

Reproducibility

The repository includes:

adapter weights and configuration;
training configuration and metrics;
experiment metadata;
aggregate evaluation results;
SHA-256 checksum for adapter weights.

The normalized training examples, split manifest, per-example predictions, full training notebook, and source pipeline are not included in this model release. Paths in the experiment metadata are relative to this repository; omitted artifacts are marked explicitly.

License

This adapter is released under the Apache License 2.0. See LICENSE for the complete license text. The Qwen3-1.7B base model is also distributed under Apache 2.0; the training datasets retain their respective licenses listed above.

Training software

Table with columns: Package, Version
Package	Version
PyTorch	2.10.0+cu128
Transformers	4.57.6
Datasets	4.8.5
Accelerate	1.14.0
BitsAndBytes	0.50.0
TRL	1.9.2
PEFT	0.20.0

Citation

FinQA

bibtex
@inproceedings{chen2021finqa,
  title={FinQA: A Dataset of Numerical Reasoning over Financial Data},
  author={Chen, Zhiyu and others},
  booktitle={Proceedings of EMNLP},
  year={2021}
}

DocFinQA

bibtex
@article{reddy2024docfinqa,
  title={DocFinQA: A Long-Context Financial Reasoning Dataset},
  author={Reddy, Varshini and others},
  year={2024}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.