AayushP418/finlora-sft-dataset

FinLoRA SFT Dataset

ChatML-formatted instruction-tuning dataset for SEC 10-K financial document question answering. Used to train finlora-sft-v2-phi35 via QLoRA supervised fine-tuning on Phi-3.5-mini-instruct.

Dataset summary

Split	Examples	Size
train	31,320	~99 MB
validation	3,479	~11 MB

Total: 34,799 examples

Sources

• EDGAR (28,548 examples) — 125 SEC 10-K annual reports across 25 S&P 500 companies, 5 sectors (Technology, Finance, Energy, Healthcare, Consumer), filings 2019–2023. Parsed from EDGAR SGML bundles, iXBRL-stripped, section-chunked at 512 tokens with 64-token overlap, MinHash LSH deduplicated (threshold=0.85, 17% removed). Answers generated by GPT-4o-mini with document context.

• FinQA (6,251 examples) — Expert-annotated numerical reasoning questions over financial tables from czyssrs/FinQA.

Format

Each record is a ChatML message list (Phi-3.5 native template):

{
  "messages": [
    {
      "role": "system",
      "content": "You are a precise financial analyst. Answer questions using ONLY the provided context. Cite sources as [SOURCE: <text fragment>]. If the answer cannot be determined from the context, say so explicitly."
    },
    {
      "role": "user",
      "content": "Context:\n<10-K passage>\n\nQuestion: <question>"
    },
    {
      "role": "assistant",
      "content": "<answer with [SOURCE: ...] citation>"
    }
  ]
}

Known limitations

• FinQA answers are bare numerical values (e.g. 0.08065) — computation-style questions may trigger the model to output raw numbers rather than prose.
• EDGAR answers were generated by GPT-4o-mini and may contain minor factual errors.
• Citation formatting in EDGAR answers is inconsistent — not all examples include a [SOURCE: ...] tag, which affects the fine-tuned model's citation quality.

Related resources

• GitHub — AayushP418/Finlora
• SFT adapter — finlora-sft-v2-phi35
• DPO adapter — finlora-dpo-v3-phi35
• Live demo