---
base_model: meta-llama/Llama-4-Scout-17B-16E-Instruct
library_name: peft
license: mit
tags:
- adaption-labs
- autoscientist
- finance
- sec-filings
- financial-arithmetic
- table-reasoning
- lora
- text-generation
- hackathon
datasets:
- narendarcodes/adaption-sec-financial-arithmetic-dataset
language:
- en
pipeline_tag: text-generation
---

# Adaption AutoScientist — SEC Financial Arithmetic (109B LoRA)

> **Powered by Adaptive Data — [Adaption Labs](https://www.adaptionlabs.ai/)**

## 🎯 Core Problem Solved

**Given a financial table from an SEC filing (10-K, 10-Q, earnings report), this model extracts the correct numeric values, identifies the right formula, executes multi-step arithmetic, and produces a verified numeric answer — never hallucinating numbers.**

Specifically, this model performs:
1. **Table-Grounded Variable Extraction** — "Given this SEC filing table with 20+ rows and 5+ columns, identify exactly which cells contain the values needed for this calculation"
2. **Multi-Step Financial Arithmetic** — "Compute the year-over-year revenue growth rate by extracting 2021 and 2022 revenue, then applying: ((2022 - 2021) / 2021) × 100"
3. **Formula-Verified Calculations** — "Every answer is cross-checked against a gold reasoning program (e.g., `subtract(8.5, 7.2) → divide(#0, 7.2) → multiply(#1, 100)`) to ensure mathematical correctness"
4. **Compound Financial Metrics** — "Compute LTI payout percentages, operating margins, debt-to-equity ratios, and other multi-variable derived values from dense financial context"

### Why This Matters
Financial analysts extract numbers from **dense corporate filings** to compute key metrics. Current LLMs:
- **Hallucinate numbers** that aren't in the table
- **Skip intermediate steps** (jumping to a final answer without showing work)
- **Apply wrong formulas** (e.g., computing growth rate with the wrong denominator)

This model is trained on **real SEC filing data** with gold reasoning programs that verify every arithmetic step, teaching it to show complete `<think>` traces before outputting a verified `<answer>`.

## Model Details

| Field | Value |
|:---|:---|
| **Trained Model Name** | `adaption_finqa_financial_reasoning` |
| **Base Model** | `meta-llama/Llama-4-Scout-17B-16E-Instruct` (109B MoE) |
| **Training Method** | Supervised Fine-Tuning (SFT) with LoRA |
| **Training Platform** | [Adaption Labs AutoScientist](https://www.adaptionlabs.ai/) |
| **Language** | English (en) |
| **License** | MIT |

## AutoScientist Platform Integration

The model was adapted using the automated training features of the Adaption Labs AutoScientist platform:

* **Adaptive Data Pipeline:** Validated the financial dataset structure, maintaining a stable Grade B (8.9 score) across multi-step calculation patterns.
* **Prompt Deduplication:** Removed redundant financial questions, focusing training on unique computation flows.
* **Prompt Rephrase Constraint:** Prompt rephrasing was disabled to protect the integrity of financial tables, column headers, and tabular structures from corruption.
* **Reasoning Traces & Blueprint Constraints:** Enforced structured arithmetic thinking. The model is trained to show complete working in a `<think>` trace verified against gold calculation programs before outputting the final answer. The blueprint constraints mandate:
  1. Extracting the exact cell coordinates and values from the input SEC table.
  2. Stating the mathematical formula explicitly.
  3. Executing arithmetic steps sequentially.
  4. Outputting a clear, verified final numeric value.
* **Hyperparameter Optimization:** Managed the training run using LoRA SFT targeting the attention and feed-forward layers of meta-llama/Llama-4-Scout-17B-16E-Instruct.
* **Evaluation:** Compared the adapted model performance against the base model on held-out financial prompts using automated preference scoring.

## 📊 Dataset

- **Dataset:** [narendarcodes/adaption-sec-financial-arithmetic-dataset](https://huggingface.co/datasets/narendarcodes/adaption-sec-financial-arithmetic-dataset)
- **Size:** 1,055 rows
- **License:** MIT

**Source Data & Attribution:**

| Source | Rows | What It Teaches | License |
|:---|:---|:---|:---|
| [czyssrs/FinQA](https://github.com/czyssrs/FinQA) | ~550 | Financial QA from real SEC filings with gold reasoning programs | MIT |
| [cerebras/TAT-QA-Arithmetic-CoT](https://huggingface.co/datasets/cerebras/TAT-QA-Arithmetic-CoT) | ~550 | Table-and-text financial QA with chain-of-thought arithmetic | Apache-2.0 |

| Metric | Before | After | Change |
|:---|:---|:---|:---|
| Grade | B | **B** | — |
| Score | 9.0 | **8.9** | -1.1% |
| Percentile | — | **28.9%** | — |

## ⚙️ Training Configuration

| Hyperparameter | Value |
|:---|:---|
| Finetune Job ID | `5213d7c9-d937-4f5b-b4f8-c7316c0a18b3` |
| Training Experiment ID | `4fa65b4d-3ed8-4189-9e8f-98a920c92a4d` |
| LoRA Rank (r) | 16 |
| LoRA Alpha | 32 |
| LoRA Dropout | 0.1 |
| Target Modules | `q_proj`, `k_proj`, `v_proj`, `o_proj`, `shared_expert.*`, `feed_forward.*` |
| Epochs | 1 |
| Batch Size | max |
| Learning Rate | 2e-5 |
| LR Scheduler | Cosine (0.5 cycles) |
| Warmup Ratio | 0.05 |
| Weight Decay | 0.05 |
| Max Grad Norm | 1.0 |

## 📈 Evaluation Results

| Evaluation Set | Base Model | Adapted Model | Winner |
|:---|:---|:---|:---|
| On Your Dataset | 44% | **58%** | ✅ Adapted |
| Across Category (Held-out) | 40% | **61%** | ✅ Adapted |

## How to Use

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-4-Scout-17B-16E-Instruct")
model = PeftModel.from_pretrained(base_model, "narendarcodes/adaption-sec-financial-arithmetic-109b")
tokenizer = AutoTokenizer.from_pretrained("narendarcodes/adaption-sec-financial-arithmetic-109b")
```

## Citation

```
@misc{golla2026secfinancial,
  title={SEC Financial Arithmetic — Table-Grounded Multi-Step Calculation from Corporate Filings},
  author={Golla Narendar},
  year={2026},
  note={Trained using Adaption Labs AutoScientist platform. Powered by Adaptive Data.}
}
```

---

> **Powered by Adaptive Data — [Adaption Labs](https://www.adaptionlabs.ai/)**
>
> Built for the [2026 Adaption AutoScientist Challenge](https://hackindia.org/2026/adaption-autoscientist-challenge-50000-prize-pool) — Finance Category

### Framework Versions
- PEFT 0.15.1