Akshat4112/finplainlm-dpo-dataset

FinPlainLM — Financial Plain-Language Preference Dataset

A preference dataset for training language models to rewrite dense financial text into clear, client-friendly plain language. Contains SFT and DPO splits ready for fine-tuning with mlx-lm-lora or any standard RLHF/DPO trainer.

Dataset Description

Split	Rows	File
DPO train	1,200	dpo/train.jsonl
DPO valid	150	dpo/valid.jsonl
SFT train	1,200	sft/train.jsonl
SFT valid	150	sft/valid.jsonl
Test sources	150	test_sources.jsonl

Data Format

DPO splits (dpo/train.jsonl, dpo/valid.jsonl)

{
  "prompt": "Rewrite the following financial text in plain, client-friendly language...",
  "chosen": "Last year, we had a profit of 9.8 million euros.",
  "rejected": "The reported operating profit for the current period stood at EUR 9.8 million..."
}

SFT splits (sft/train.jsonl, sft/valid.jsonl)

{
  "messages": [
    {"role": "user", "content": "Rewrite the following financial text..."},
    {"role": "assistant", "content": "Last year, we had a profit of 9.8 million euros."}
  ]
}

Test sources (test_sources.jsonl)

{"source": "Operating profit was EUR 9.8 mn..."}

Construction

Source sentences: Financial PhraseBank (sentences_allagree config, 2,246 sentences deduped to 1,500 substantive ones ≥6 words).

Teacher model: mlx-community/Qwen2.5-14B-Instruct-4bit running locally on Apple Silicon.

Chosen rewrite: Best-of-3 candidates selected by a composite score: 100 × faithful_bonus − flesch_grade − 2 × jargon_count, where:

• faithful_bonus = 1 if all numbers and named entities from the source are preserved
• flesch_grade = Flesch-Kincaid grade level (lower = simpler)
• jargon_count = count of finance-domain jargon terms

Rejected rewrite: One of four degradation types cycled across examples:

• over_formal — rewritten with formal register and longer sentences
• jargon_padded — original jargon terms preserved or added
• literal — near-literal restatement without simplification
• verbose — correct content but padded with unnecessary words

Split: 80 / 10 / 10 (train / valid / test), shuffled with seed 0.

Intended Use

Fine-tune a 7B instruction model (e.g. Qwen2.5-7B-Instruct) on the SFT split first, then apply DPO with the SFT model as frozen reference. This trains the model to prefer simpler rewrites without sacrificing factual accuracy.

The companion fine-tuned model is available at Akshat4112/FinPlainLM.

Limitations

• Domain narrow: Source sentences are from Financial PhraseBank, which covers short news-wire snippets. Performance on longer documents, earnings calls, or regulatory filings has not been evaluated.
• Teacher model bias: Chosen rewrites were generated by Qwen2.5-14B-Instruct. Any stylistic biases of that model are inherited by this dataset.
• Rejected quality: Rejected rewrites are synthetically degraded, not naturally occurring bad rewrites. Real-world bad rewrites may differ in character.
• English only: All sentences are in English. The model trained on this data should not be expected to generalize to other languages.
• Non-commercial only: CC BY-NC-SA 3.0 license prohibits commercial use.

Ethical Considerations

• Factual accuracy: The chosen rewrite heuristic penalises loss of numbers and named entities, but does not guarantee factual correctness. Always verify model outputs before using in client-facing communications.
• Financial advice: This dataset is for text simplification only. Nothing in it constitutes financial advice, and models trained on it should not be deployed as financial advisors.
• Source license compliance: Financial PhraseBank is CC BY-NC-SA 3.0. This dataset inherits that license; any derivative work must do the same.

License

CC BY-NC-SA 3.0 — inherited from Financial PhraseBank (Malo et al., 2014). Non-commercial use only. Derivatives must carry the same license.

Citation

If you use this dataset, please cite the original Financial PhraseBank:

@article{malo2014good,
  title={Good debt or bad debt: Detecting semantic orientations in economic texts},
  author={Malo, Pekka and Sinha, Ankur and Korhonen, Pekka and Wallenius, Jyrki and Takala, Pyry},
  journal={Journal of the Association for Information Science and Technology},
  volume={65},
  number={4},
  pages={782--796},
  year={2014}
}