| --- |
| base_model: unsloth/gemma-4-31b-it-unsloth-bnb-4bit |
| library_name: peft |
| pipeline_tag: text-generation |
| tags: |
| - base_model:adapter:unsloth/gemma-4-31b-it-unsloth-bnb-4bit |
| - grpo |
| - lora |
| - sft |
| - transformers |
| - trl |
| - unsloth |
| license: cc-by-4.0 |
| datasets: |
| - mmmikolajczak/st312-data |
| language: |
| - en |
| model-index: |
| - name: FinLM-Reasoning |
| results: |
| - task: |
| type: text-generation |
| name: Financial numerical reasoning |
| dataset: |
| name: FinTest |
| type: fintest |
| metrics: |
| - name: All tasks |
| type: aggregate_score |
| value: 48.55 |
| - name: Held-out tasks |
| type: held_out_aggregate_score |
| value: 54.95 |
| - name: Numerical reasoning |
| type: numerical_reasoning_aggregate_score |
| value: 78.00 |
| --- |
| |
| # FinLM-Reasoning |
|
|
| **FinLM-Reasoning is an open financial reasoning model built to bring frontier-level numerical finance performance to teams that need accuracy, transparency, and low inference cost.** |
|
|
| Financial workflows are moving from keyword search and brittle pipelines to AI agents that can read filings, reason over tables, perform multi-step calculations, and answer domain-specific financial questions. FinLM-Reasoning is designed for that future: a finance-specialized reasoning model optimized for numerical QA, table reasoning, and calculation-heavy financial workflows. |
|
|
| ## Models |
|
|
| | Model | Link | Best for | |
| |---|---|---| |
| | FinLM-31B | https://huggingface.co/marco-molinari/FinLM-31B | General financial NLP, financial QA, filing analysis, summarization, classification, and broad finance workflows | |
| | FinLM-Reasoning | https://huggingface.co/marco-molinari/FinLM-Reasoning | Numerical financial reasoning, table QA, multi-step calculations, FinQA, ConvFinQA, and TAT-QA-style tasks | |
|
|
| ## Why FinLM-Reasoning |
|
|
| - **Frontier-level financial reasoning:** FinLM-Reasoning reaches the top score on the FinTest numerical reasoning aggregate among evaluated models. |
| - **Built for calculation-heavy finance tasks:** Optimized for numerical QA, table reasoning, financial statement analysis, and multi-step financial calculations. |
| - **GRPO-optimized checkpoint:** FinLM-Reasoning is trained from FinLM-31B and further optimized for numerical financial reasoning with GRPO. |
| - **Open and practical:** Designed to offer strong reasoning performance at materially lower inference cost than large closed-source frontier systems. |
| - **Multimodal-ready foundation:** Built from a multimodal base model and validated on extractive QA over rendered SEC filings. |
|
|
| ## Headline results |
|
|
| | Model | All FinTest tasks | Held-out tasks | Financial reasoning | |
| |---|---:|---:|---:| |
| | FinLM-31B | 52.79 | 65.94 | 65.00 | |
| | FinLM-Reasoning | 48.55 | 54.95 | 78.00 | |
| | Gemma-4-31B base | 43.79 | 46.44 | 41.33 | |
| | GPT-5.5 | 52.31 | 54.24 | 66.78 | |
| | Gemini 3 Flash | 53.91 | 60.63 | 56.18 | |
| | Claude Opus 4.7 | 55.97 | 59.91 | 72.80 | |
|
|
| Scores are percentages. Numerical reasoning aggregates FinQA, TAT-QA, and ConvFinQA. |
|
|
| ## When to use which model |
|
|
| Use **FinLM-31B** when you need the strongest general financial model across a wide range of tasks. |
|
|
| Use **FinLM-Reasoning** when your workload is calculation-heavy, such as financial table QA, multi-step numerical questions, or reasoning over earnings reports. |
|
|
| ## Example use cases |
|
|
| - Financial table QA |
| - Multi-step financial calculations |
| - Numerical reasoning over earnings reports |
| - Filing and earnings-call analysis |
| - Financial question answering |
| - KPI and metric extraction |
| - Analyst workflow automation |
| - Financial research assistants |
| - Risk and compliance research workflows |
| - Benchmarking financial AI systems |
|
|
| ## Responsible use |
|
|
| FinLM models are research and development tools. They should not be used as the sole basis for investment decisions, credit decisions, lending decisions, insurance decisions, legal conclusions, regulatory filings, fraud accusations, or consumer financial advice. |
|
|
| For production use, we recommend human review, audit logs, domain-specific evaluation, bias testing, and validation against authoritative financial sources. |
|
|
| ## Citation |
| ``` |
| |
| @misc{finlm_suite_2026, |
| title = {The FinLM Suite: Pushing the Limits of Open Source Financial Language Modeling}, |
| author = {Marco Molinari, Mateusz Mikolajczak, Luca Imeneo, Alvise Sembenico, Deasy Darlene Tunas, Saharsha Navani}, |
| year = {2026}, |
| } |
| |
| ``` |
