ramtinasadi/Quantfinlab-Qwen3.5-2B-Financial-Analysis-GGUF

Qwen3.5-2B Quantfinlab Financial Analyst GGUF

Qwen3.5-2B Quantfinlab Financial Analyst is an English financial-analysis adaptation of Qwen/Qwen3.5-2B, developed for Financial Analysis with Local LLM in the Quantitative Finance Lab.

This repository contains the Q4_K_M GGUF used for local inference in Project 24. The model was adapted for evidence-grounded financial analysis rather than generic financial question answering. Its intended tasks include SEC filing analysis, financial-event analysis, macro-release interpretation, evidence reconciliation, and market/cross-asset synthesis.

The model is designed to work with the Financial Analysis with Local LLM pipeline, where current documents and quantitative context are retrieved or calculated outside the language model and then supplied as evidence.

Project and reproducibility links

• —Financial Analysis with Local LLM
• —Quantitative Finance Lab: github.com/ramtin-asadi/Quantitative-Finance-Lab
• —Full training notebook: qwen_train_lora.ipynb
• —Open training notebook in Colab: Google Colab
• —LoRA adapter: Quantfinlab-Qwen3.5-2B-Financial-Analysis-LoRA
• —Base model: Qwen/Qwen3.5-2B

What the model is capable of

The supervised adaptation targets five related financial-analysis task families:

• —Financial-event analysis: identify what happened, what changed, its materiality, why it may matter, and which evidence supports the conclusion.
• —SEC filing-change analysis: compare disclosures across filings and identify material changes in risk factors, liquidity, debt, operating performance, capital spending, contingencies, and other company disclosures.
• —Macro-release analysis: interpret official macroeconomic releases in the context of prior values, revisions, recent trends, and supplied market context.
• —Evidence reconciliation: combine multiple source excerpts while distinguishing direct facts, interpretation, and uncertainty.
• —Market and cross-asset synthesis: connect supplied market, rates, risk, volatility, credit, macro, and cross-asset measures without treating correlation as proof of causality.

The adapted model is trained to return structured analysis containing a conclusion, materiality assessment, evidence-linked claims, what changed, why it matters, and explicit uncertainty.

How we use this model in Financial Analysis with Local LLM

We separate deterministic financial computation from language-model synthesis.

The workflow combines:

1. 1.Primary documents and official releases, including SEC filings and U.S. macro/policy sources.
2. 2.Point-in-time filtering and lexical retrieval, so evidence must have been available by the requested analysis cutoff.
3. 3.Calculated Quantfinlab context, including selected measures from earlier Quantfinlab projects covering market risk, volatility, the yield curve, financial conditions, factors, cross-asset state, fundamentals, credit, and macro conditions.
4. 4.Qwen3.5-2B Quantfinlab Financial Analyst, which interprets and synthesizes the supplied evidence.
5. 5.Post-generation validation, which checks structured output, evidence references, numerical traceability, timestamps, and unsupported claims before an answer is accepted.

The model is not expected to memorize current financial data. Current facts are supplied at runtime by the surrounding analysis system.

For an end-to-end example, see Project 24.

Training

The model was produced from a pinned revision of Qwen3.5-2B using a text-focused 16-bit LoRA workflow.

LoRA was applied to the language-model attention and MLP projections:

q_proj
k_proj
v_proj
o_proj
gate_proj
up_proj
down_proj

Training used a reviewed, frozen train/validation dataset with point-in-time evidence checks. Validation was strictly later than training, and the training notebook checks for leakage across example IDs, event/group IDs, source IDs, and source hashes.

The full training workflow is available in qwen_train_lora.ipynb.

Export and quantization

After LoRA training:

1. 1.the adapter was checked on held-out examples;
2. 2.the adapter was merged with the exact pinned Qwen base revision;
3. 3.the merged model was converted to F16 GGUF;
4. 4.a pinned llama.cpp revision was used for conversion and quantization;
5. 5.the final model was quantized to Q4_K_M;
6. 6.the exported GGUF was loaded through llama.cpp and checked again after quantization;
7. 7.SHA256 checksums, the chat-template hash, export metadata, and the dataset manifest were recorded.

The GGUF export is therefore tied to the specific trained adapter and base-model revision rather than being an unrelated third-party quantization.

Acceptance checks

The published adapter passed 30/30 held-out acceptance checks on the first attempt: six validation examples from each of the five task families.

The final Q4KM GGUF also passed 5/5 post-export task checks.

The checks cover:

• —structured-output/schema validity;
• —evidence-ID validity;
• —numerical traceability to supplied evidence;
• —normal generation termination;
• —repeated-claim detection;
• —compatibility of the quantized model with the expected prompt/template behavior.

These checks are release/grounding checks. They do not constitute a benchmark of financial judgment, forecasting accuracy, investment performance, or superiority over other language models.

Download

Using the Hugging Face CLI:

hf download ramtinasadi/Quantfinlab-Qwen3.5-2B-Financial-Analysis-GGUF \
  Quantfinlab-Qwen3.5-2B-Financial-Analysis-Q4_K_M.gguf

Intended use

This model is intended for:

• —local financial-analysis experiments;
• —evidence-grounded analysis of supplied financial documents;
• —SEC filing comparison and disclosure analysis;
• —interpretation of supplied macroeconomic releases;
• —financial-event analysis with evidence references;
• —market and cross-asset synthesis when structured context is supplied;
• —reproducible experimentation with a small local financial SLM.

Limitations

This is a 2B-parameter model and should be used with retrieval, deterministic calculations, and validation for serious financial analysis.

Important limitations include:

• —The model does not contain live market data.
• —It can produce incorrect interpretations even when the supplied facts are correct.
• —Evidence grounding reduces hallucination risk but does not eliminate it.
• —Causal interpretations of market moves are particularly uncertain and should be reviewed.
• —The model was adapted for English financial analysis.
• —The fine-tuning focused on text; vision capabilities were not fine-tuned.
• —Outputs are not investment recommendations and should not replace independent financial verification.

License and attribution

This model is a derivative of Qwen/Qwen3.5-2B, released under the Apache License 2.0. This repository uses the same license.

Please retain attribution to the original Qwen model when redistributing derivative artifacts.

Author

Developed by Ramtin Asadi as part of the open-source Quantitative Finance Lab.