---
base_model: google/gemma-4-e2b-it
library_name: peft
pipeline_tag: text-generation
tags:
- base_model:adapter:google/gemma-4-e2b-it
- lora
- peft
- transformers
- unsloth
- finance
- sec-filings
- xbrl
---

# FinEdgar Gemma 4 E2B LoRA

This repository contains a LoRA adapter fine-tuned for question answering over SEC filings.

The adapter is intended to be used inside a retrieval-augmented financial QA system. It was trained to answer with filing-grounded context and to work alongside deterministic XBRL tooling for structured financial facts.

## Base Model

- Base model: `google/gemma-4-e2b-it`
- Adapter type: LoRA
- PEFT version: `0.18.1`
- Task type: causal language modeling

Users must have access to the base model and comply with the base model license and terms.

## Intended Use

This adapter is intended for:

- SEC filing question answering
- filing-grounded financial summaries
- structured financial QA when paired with XBRL facts
- local RAG pipelines over 10-K, 10-Q, and 8-K filings

It is not intended to be used as a standalone source of financial truth. Numeric answers should be checked against SEC XBRL facts or the original filing.

## Out-of-Scope Use

Do not use this model as financial advice, investment advice, accounting advice, legal advice, or a substitute for reviewing original SEC filings.

The model can be wrong when retrieval context is missing, incomplete, stale, or ambiguous.

## Training Data

Training used a mixture of:

- financial QA examples
- financial sentiment examples
- SEC/XBRL-derived synthetic QA
- filing-diff summary examples

FinanceBench was used for evaluation and was not used as direct training supervision.

## Evaluation

Latest tracked FinanceBench result from the project evaluation run:

| Metric | Result |
|---|---:|
| Overall accuracy | 40.7% |
| XBRL route accuracy | 40.0% |
| RAG route accuracy | 41.1% |
| Recall@5 | 30.0% |
| Average latency | 2,144 ms |

Evaluation file: `eval_latest_model_20260414.json`

These results are system-level results from the full FinEdgar pipeline, not adapter-only generation in isolation.

## Limitations

- The adapter depends on good retrieval context.
- It can produce incomplete narrative answers when evidence is spread across multiple filings or sections.
- It may format or reason about financial metrics incorrectly without deterministic XBRL support.
- It does not know current market prices or events unless those are provided in context.
- It should be used with citations and source checking.

## Loading

Example PEFT loading pattern:

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

base_model = "google/gemma-4-e2b-it"
adapter = "YOUR_USERNAME/finedgar-gemma-4-e2b-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, adapter)
```

Depending on the installed Transformers version and Gemma 4 support, a model-specific class or processor may be required instead of `AutoModelForCausalLM`.

## Suggested Inference Pattern

Use the model with retrieved filing context:

```text
Question: <user question>

Context:
<retrieved SEC filing excerpts with citations>

Answer using only the provided context. If the answer is not present, say so.
```

For structured metrics, prefer XBRL lookup and use generation only for explanation or formatting.