dipanjann

Gemma-4-31b-it-FinQA

README

License: apache-2.0

Model Details

Base model: google/gemma-4-31B-it
Adapter type: LoRA / PEFT
PEFT type: LORA
LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0.05
Target modules: q_proj, v_proj
Adapter weights: adapter_model.safetensors
Source archive checksum: 66b80c699208a913b502534eeff46f132cbd6f0b3b610dde5a16fc814c5fe8bf

Intended Use

Use this adapter for research and prototyping on financial QA prompts that include a question plus relevant financial-report context. Do not use it as a substitute for professional financial advice, audited filing review, or production decisioning without independent validation.

Training Data

This adapter is intended for the FinQA training workflow in the FinAI Dexlabs repository:

Processed file: data/processed/finqa/train_question_answer.jsonl
Records: 6,251
Processed checksum: 58244847a8b9958a98260d843b5fc81a7cd74fce39fe2f1e2a6ba9ec50fb5e38
Fields retained: question, answer, pre_text, table, table_ori, post_text, steps

Dataset readiness note: the FinQA export passes required completeness checks, but the quality report flags duplicate primary questions. Treat this model as a research adapter until model-level evaluation and split-hygiene evidence are added.

Evaluation

No model-level evaluation results are included in this release. Add scores only after a reproducible evaluation run exists, including dataset, metric, command, run date, and exact score.

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "google/gemma-4-31B-it"
adapter_id = "dipanjann/Gemma-4-31b-FinQA"

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

prompt = """Answer the financial question using the context.

Question: What was the change in revenue?

Context: <paste relevant financial report context here>
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

This is an adapter, not a standalone merged full-weight model.
Users must load it with the compatible Gemma 4 31B base model.
Financial numeric answers should be independently checked.
Dataset-level readiness is not the same as model-level evaluation.
Duplicate or near-duplicate source questions can affect benchmark interpretation if split hygiene is not enforced.

Expected Files

README.md
adapter_config.json
adapter_model.safetensors
chat_template.jinja
config.json
special_tokens_map.json
tokenizer.json
tokenizer_config.json
trainer_state.json

Available on FriendliAI

Dedicated Endpoints

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

Learn more

Model Details

Model Provider

dipanjann

Model Tree

Base

google/gemma-4-31B-it

Adapter

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model Details

Base model: google/gemma-4-31B-it
Adapter type: LoRA / PEFT
PEFT type: LORA
LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0.05
Target modules: q_proj, v_proj
Adapter weights: adapter_model.safetensors
Source archive checksum: 66b80c699208a913b502534eeff46f132cbd6f0b3b610dde5a16fc814c5fe8bf

Intended Use

Training Data

This adapter is intended for the FinQA training workflow in the FinAI Dexlabs repository:

Processed file: data/processed/finqa/train_question_answer.jsonl
Records: 6,251
Processed checksum: 58244847a8b9958a98260d843b5fc81a7cd74fce39fe2f1e2a6ba9ec50fb5e38
Fields retained: question, answer, pre_text, table, table_ori, post_text, steps

Evaluation

No model-level evaluation results are included in this release. Add scores only after a reproducible evaluation run exists, including dataset, metric, command, run date, and exact score.

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = "google/gemma-4-31B-it"
adapter_id = "dipanjann/Gemma-4-31b-FinQA"

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

prompt = """Answer the financial question using the context.

Question: What was the change in revenue?

Context: <paste relevant financial report context here>
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

This is an adapter, not a standalone merged full-weight model.
Users must load it with the compatible Gemma 4 31B base model.
Financial numeric answers should be independently checked.
Dataset-level readiness is not the same as model-level evaluation.
Duplicate or near-duplicate source questions can affect benchmark interpretation if split hygiene is not enforced.

Expected Files

README.md
adapter_config.json
adapter_model.safetensors
chat_template.jinja
config.json
special_tokens_map.json
tokenizer.json
tokenizer_config.json
trainer_state.json