Downloads · 30 days
123
5% of all-time downloads
dschauhan08/Qwen2.5-3B-Finance-GGUF
Qwen2.5-3B-Finance-GGUF is a machine learning model from dschauhan08. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Model overview Name: SLM-Finance (Qwen2.5-3B-based small finance language model) Architecture: 3B-parameter Qwen2.5-style decoder-only transformer, instruction-tuned and then further fine-tuned for multi-step financia…
Downloads · 30 days
123
5% of all-time downloads
All-time downloads
2.7K
Public
Repo size
8.1 GB
Likes
1
Public
Click a slice to open those files.
.gguf8.1 GB · 100%
From the Hugging Face model README
Model overview Name: SLM-Finance (Qwen2.5-3B-based small finance language model) Architecture: 3B-parameter Qwen2.5-style decoder-only transformer, instruction-tuned and then further fine-tuned for multi-step financial reasoning. This model card follows the spirit of standardized model reporting practices proposed for model cards in the ML literature.
The model is optimized for “thinking-style” analysis of equities, portfolios, and risk metrics rather than general conversation. It is designed to produce structured reasoning steps followed by concise decisions (for example, Buy/Sell/Hold plus explanation).
Intended use Primary intended uses
Educational and research use for learning about quantitative trading concepts (Sharpe, Sortino, volatility estimators, risk/return, portfolio optimization).
Prototyping trading logic, scenario analysis, and strategy explanation in backtesting or paper-trading environments.
Assisting with interpretation of time‑series market data, simple portfolio analytics, and high-level readings of public company disclosures (e.g., SEC 10‑K/10‑Q risk sections).
Out of scope / not for
Live execution of trades, production portfolio management, or any use where model errors can cause real financial loss.
Providing investment advice, personalized financial planning, or tax/regulatory guidance.
Predicting future prices or guaranteeing outperformance versus benchmarks such as the S&P 500 .
Training data and process The base model is a Qwen2.5-3B Instruct checkpoint that was already instruction-tuned on general dialogue and coding tasks, then further fine-tuned on finance‑specific data. This second-stage fine‑tuning mixes multiple custom datasets covering: synthetic trading journals, price/volume tables (OHLCV), simple options-like structures, and formatted conversations where the assistant explains trading decisions and portfolio risk.
The fine-tuning objective is supervised learning: the model is trained to produce both intermediate “thinking” (rationale) and final answers in a stable, machine‑readable format. No proprietary broker data, private customer records, or non‑public information are included; all real-world ingredients (e.g., SEC filings or historical prices) should be sourced from public datasets or synthetic generators you control.
Inputs, outputs and formats Inputs
Natural language questions about: single‑name equities, baskets, risk metrics, portfolio construction, and strategy comparisons.
Tabular or semi-structured snippets pasted into the prompt (e.g., recent OHLCV rows, portfolio weights, factor betas, or extracted parts of a 10‑K risk section).
Outputs
A mix of:
Explanatory reasoning text (multi-step “thinking” about drivers, risks, and trade‑offs).
A final, clearly delimited answer section (e.g., JSON with fields like signal, horizon, confidence, rationale_short).
The recommended prompting pattern is “system = hedge‑fund PM / quant”, “user = provides data + question”, and “assistant = outputs reasoning then a concise, structured decision block.
Limitations, risks and ethical considerations This model is not a registered investment adviser and does not have access to live markets, order books, or confidential company information, so its views are at best approximate and often outdated or wrong. It may confidently produce incorrect calculations, misinterpret regime changes, or overgeneralize from synthetic distributions that do not match real‑world market conditions.
Because training emphasizes reasoning style, the model may sound persuasive even when its trade recommendations would be unacceptable under real‑world risk, liquidity, or compliance constraints. Users must never use its outputs as the sole basis for trading decisions, risk management, or client communication, and should add guardrails such as position limits, independent backtests, and human review.
Evaluation and monitoring You should document and periodically re‑run evaluations on:
Numerical correctness: Sharpe/Sortino/volatility calculations on held‑out synthetic and real price series.
Decision stability: Consistency of signals (e.g., Buy vs Hold) under small prompt perturbations and minor data changes.
Domain coverage: Performance across single‑day scalping scenarios, swing/multi‑day strategies, long‑horizon portfolio allocation, and SEC‑style risk extraction.
All evaluations must be considered approximations; before any real‑money deployment, measure the model’s behavior in a sandboxed research environment with your own constraints, instruments, and compliance rules.