Model Card: Financial Reasoning QA — Llama 3.3 70B LoRA

Model Details

Model Description

This is a LoRA adapter fine-tuned on top of Llama-3.3-70B-Instruct to do step-by-step numeric reasoning over personal finance questions. It covers both generic/global reasoning tasks (debt payoff strategy comparison, loan amortization, retirement projection, budget trade-offs, opportunity-cost decisions between paying off debt vs. investing, emergency-fund trade-offs) and Kenya/East-Africa-specific financial reasoning (KRA PAYE take-home pay calculations including NSSF/Housing Levy/SHIF, SACCO vs. bank loan comparisons, chama rotating-savings timing, mobile-money savings products).

Every training answer shows worked, step-by-step arithmetic rather than a bare final number — training labels for the generic and Kenya-localized tasks were computed programmatically (Python) rather than hand-written, so ground-truth correctness of the underlying numbers is guaranteed by construction.

Submitted to the AutoScientist Challenge (Personal Finance track) by Adaption Labs.

Developed by: Grold Otieno Mboya
Shared by: Grold Otieno Mboya
Model type: LoRA adapter for causal language modeling (personal finance numeric reasoning)
Language(s): English
License: Llama 3.3 Community License Agreement (inherited from the base model — see https://www.llama.com/llama3_3/license/)
Finetuned from model: togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference

Model Sources

Dataset: https://huggingface.co/datasets/Gro97/financial_reasoning_qa
Dataset (Kaggle): https://www.kaggle.com/datasets/groldotieno/adaption-financial-reasoning-qa
Model (Kaggle): https://www.kaggle.com/models/groldotieno/financial-reasoning-qa-llama3-3-70b-lora
Demo interface: https://kenya-finance-math.adaptionlabs.app/
Challenge: AutoScientist Challenge, Adaption Labs

Uses

Direct Use

Answering personal-finance numeric-reasoning questions with worked, step-by-step arithmetic: comparing debt payoff strategies, projecting loan or retirement balances, weighing budget trade-offs, deciding between paying down debt vs. investing, and reasoning through Kenya-specific financial products (SACCO vs. bank loans, chama contribution timing, mobile-money savings, KRA PAYE take-home pay).

Downstream Use

Could serve as a starting point for further fine-tuning on adjacent personal-finance or localized-financial-literacy tasks, or as a component in an educational finance tool aimed at Kenyan/East African users.

Out-of-Scope Use

Not intended as a substitute for licensed financial or tax advice. Not evaluated for general capability, factuality, or safety beyond the Personal Finance task category described here. Illustrative/generic tax brackets used in the non-localized tax-reasoning task type are not tied to any specific country's real tax code and should not be treated as such.

Bias, Risks, and Limitations

The training data is synthetic and procedurally generated across the task types described above; the Kenya-localized content (NSSF, Housing Levy, SHIF, PAYE bands) reflects rates believed accurate as of the dataset's construction and may drift out of date if those rates change.

Known arithmetic limitation: direct testing against DeepSeek on identical loan-amortization questions found this model's generated arithmetic can be off by a small margin on multi-step calculations involving exponentiation (e.g. (1+r)^n in reducing-balance loan formulas) — a general limitation of LLM-generated arithmetic (numbers are generated token-by-token, not computed), not specific to this fine-tune. This affects single-formula questions, not just the more obviously iterative ones (e.g. month-by-month debt payoff simulations). For exact figures, pair this model with an actual calculator/code-execution step rather than trusting its generated arithmetic directly — the deployed demo interface includes a disclaimer to this effect for the same reason.

Recommendations

Users should independently verify any numeric output used for real financial decisions, and should not assume performance on Kenya-specific financial products generalizes to other countries' financial systems.

How to Get Started with the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference"
adapter_id = "Gro97/financial-reasoning-qa-llama3-3-70b-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto")
model = PeftModel.from_pretrained(base_model, adapter_id)

prompt = "KES 300,000 SACCO loan at 1% per month reducing balance over 24 months — total cost?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=400)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Training Details

Training Data

Custom-built dataset financial_reasoning_qa: 3,456 rows (prompt/input/output triples) across thirteen task types, mixed approximately as follows:

Task type	Share
Debt payoff strategy (avalanche vs. snowball)	19%
SACCO vs. bank loan comparison	10%
Retirement/compound-growth projection	9%
Loan amortization — extra-payment impact	9%
Opportunity cost (pay off debt vs. invest)	8%
Chama rotating-savings timing	8%
KRA PAYE take-home pay	7%
SACCO/mobile-money savings goal timing	6%
Raise allocation (pay debt vs. invest, Kenya)	6%
Budget trade-off / savings-goal timing	5%
Emergency-fund trade-off	4%
KRA PAYE raise impact	4%
Generic illustrative tax-bracket reasoning	2%

Every row's numeric answer was computed programmatically (Python) rather than hand-written. Full generator script and version history (including earlier rejected versions — see Limitations of the dataset itself for that history) are in the dataset repository.

Hugging Face: https://huggingface.co/datasets/Gro97/financial_reasoning_qa
Kaggle: https://www.kaggle.com/datasets/groldotieno/adaption-financial-reasoning-qa

Training Procedure

Fine-tuned via Adaption Labs' AutoScientist platform using LoRA.

Training Hyperparameters

Training regime: see adapter_config.json in this repository for the exact LoRA rank/alpha/target-module configuration used for this run.

Speeds, Sizes, Times

Not independently benchmarked by the author; training was run on Adaption Labs' hosted infrastructure (specific hardware not disclosed to the end user for this run).

Evaluation

Testing Data, Factors & Metrics

Testing Data

Two evaluation sets, as reported by Adaption Labs' AutoScientist platform: (1) a held-out split of this project's own dataset, (2) a broader set of unseen tasks from Adaption's internal Personal Finance category benchmark.

Metrics

Pairwise preference win rate against the un-adapted base model, as reported by Adaption Labs' AutoScientist platform (judge methodology not disclosed to the end user).

Results

Evaluation set	Base model win rate	Adapted model win rate
This dataset's held-out samples	10%	90%
Broader Personal Finance category	11%	89%

Summary

The adapter shows a large improvement over the base model both on its own training distribution (90% win rate) and on the broader Personal Finance category benchmark (89% win rate) — the two numbers are close, suggesting the gains generalize reasonably well beyond the specific training distribution, unlike cases where own-dataset performance substantially outpaces category-wide performance.

Environmental Impact

Hardware Type: Not disclosed by the training platform for this run.
Hours used: Not disclosed.
Cloud Provider: Adaption Labs' hosted infrastructure.
Compute Region: Not disclosed.
Carbon Emitted: Not calculated.

Technical Specifications

Model Architecture and Objective

LoRA adapter applied to togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference, trained via supervised fine-tuning to improve step-by-step numeric reasoning over personal finance questions.

Compute Infrastructure

Software

PEFT 0.15.1
Trained via Adaption Labs' AutoScientist platform

Citation

BibTeX:

@misc{mboya2026financialreasoning,
  author = {Mboya, Grold Otieno},
  title = {Financial Reasoning QA: A LoRA Adapter for Personal Finance
           Numeric Reasoning with Kenya/East-Africa Localization},
  year = {2026},
  howpublished = {AutoScientist Challenge submission, Adaption Labs},
  url = {https://huggingface.co/Gro97/financial-reasoning-qa-llama3-3-70b-lora}
}

Model Card Contact

See author's Hugging Face profile: https://huggingface.co/Gro97

Framework versions

PEFT 0.15.1

Downloads last month: 20

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Gro97/financial-reasoning-qa-llama3-3-70b-lora

Base model

togethercomputer/Meta-Llama-3.3-70B-Instruct-Reference

Adapter

(878)

this model