stindardlogic/debate-argumentation-sft-100k

Debate & Argumentation SFT (100K)

100,000 ShareGPT conversations covering persuasive writing, steel-manning, rebuttal, policy analysis, and Socratic dialogue. Each example trains models to construct well-structured arguments, anticipate counterarguments, and engage in rigorous intellectual discourse.

Motivation

A persistent gap in LLM capabilities is the ability to reason and argue well — not just describe positions, but construct arguments with premises, evidence, and logical structure. Models trained primarily on factual Q&A tend to:

• Present both sides of an issue without committing to a position when asked for a persuasive argument
• Fail to steelman opposing views (describe them in their weakest form)
• Produce rebuttals that miss the actual claim being made
• Write policy arguments that lack specificity about mechanisms and tradeoffs

This dataset trains models to engage with controversial topics at a higher intellectual level: arguing a position persuasively when asked, charitably representing opposing views, and providing substantive policy analysis.

Dataset Description

100,000 conversations across 7 task types and 17 topics:

Task Type Distribution

Task Type	Count	%	Description
persuasive_essay	~31,250	31%	Argue a position with premises, evidence, and conclusion
steel_man	~18,750	19%	Present the strongest version of an opposing view
rebuttal	~18,750	19%	Directly counter a specific argument
counter_argument	~12,500	13%	Argue against a commonly assumed position
policy_memo	~6,250	6%	Professional policy analysis with recommendations
socratic	~6,250	6%	Socratic dialogue that uncovers tensions and distinctions
debate_opener	~6,250	6%	Structured opening statement for a formal debate

Topics Covered

Workplace & Economics:

• Permanent remote work options
• Four-day workweek policy
• Corporate sabbatical programs
• Meeting length defaults (25-minute default)
• Universal Basic Income

Technology & Policy:

• AI regulation (pharmaceutical-style approval)
• Open source AI development
• Germline gene editing
• Electric vehicle mandate pace
• Data privacy legislation

Society & Values:

• Social media bans for under-16
• Meritocracy critique
• Surveillance and privacy (nothing-to-hide argument)
• Justice (equal outcomes vs. equal opportunity)

Science & Energy:

• Nuclear power for decarbonization
• AI creativity and authorship

Format

{
  "conversations": [
    {
      "from": "human",
      "value": "Write a persuasive essay arguing that companies should offer permanent remote work options for all knowledge workers. Make a strong case — I want to be convinced."
    },
    {
      "from": "gpt",
      "value": "The case for permanent remote work rests on three pillars that, taken together, are difficult to dismiss: demonstrated productivity, talent market realities, and equity...

[full persuasive argument with premises, evidence, and conclusion]"
    }
  ],
  "metadata": {
    "task_type": "persuasive_essay",
    "topic": "remote_work"
  },
  "id": "abc123"
}

Key Properties of Responses

1. Position commitment: When asked to argue a position, responses argue it — they don't hedge with "on the other hand" or present a false balance. A persuasive essay is persuasive.

2. Argument structure: Responses identify premises, provide evidence or reasoning for each, and build to a conclusion. Not just assertions.

3. Charitable representation (steel-manning): Steel-man responses present the opposing view in its strongest form — the version that proponents would recognize and endorse, not a strawman.

4. Targeted rebuttals: Rebuttal responses identify the specific claim being rebutted and address the actual argument, not a weaker version of it.

5. Intellectual rigor in dialogue: Socratic dialogues surface genuine tensions, make distinctions, and avoid strawmanning either position in the dialogue.

6. Policy specificity: Policy memos include mechanisms, tradeoffs, and implementation considerations — not just values statements.

Use Cases

• SFT fine-tuning for improved argumentation and persuasive writing
• Training AI debate assistants and writing aids
• Improving model performance on argument quality benchmarks
• Building AI tutors for rhetoric and critical thinking education
• Complementary to alignment datasets: models should argue positions when asked, not just describe them neutrally
• Paired with sycophancy-reduction-dpo-100k for models that argue well but don't cave to pushback

License

Apache 2.0