dataset_info:
features:
- name: messages
list:- name: role
dtype: string - name: content
dtype: string
splits:
- name: role
- name: train
num_bytes: 13130669
num_examples: 2672 - name: validation
num_bytes: 1461857
num_examples: 296
download_size: 14607776
dataset_size: 14592526
configs: - config_name: default
data_files:- split: train
path: data/train-* - split: validation
path: data/validation-*
- split: train
Critical Interlocutor — SFT dataset
Supervised fine-tuning data for turning a small open model (target: Qwen3 1.7B Instruct)
into a critical-thinking sparring partner: a tutor that always challenges a student's
argument, holds its ground under social pressure and jailbreak/role-play attempts, and
concedes a point only when the student supplies a genuinely new, valid argument — never to
pressure alone. It challenges interpretations and reasoning, never settled facts.
Format
OpenAI/HF chat messages format — a system prompt (the interlocutor instructions) followed
by alternating user (student) and assistant (tutor) turns.
{"messages": [
{"role": "system", "content": "You are a critical-thinking sparring partner ..."},
{"role": "user", "content": "<student claim>"},
{"role": "assistant", "content": "<tutor challenge>"},
{"role": "user", "content": "<pressure / jailbreak / new argument>"},
{"role": "assistant", "content": "<holds, or concedes only to a real argument>"}
]}
Size
| Split | Examples |
|---|---|
| train | 1,846 |
| validation | 205 |
| total | 2,051 |
Dialogues are 1–9 turns (2–19 messages). ~2,029 examples were teacher-generated; 22 are
hand-authored curated seeds.
How it was built
- Teacher distillation. A frontier model (
gpt-4.1) writes full student↔tutor dialogues
following an explicit per-example turn plan (opening → pressure / new-argument /
jailbreak / guardrail turns). - Diversity sampling. Each example is sampled independently across orthogonal axes so the
model learns the disposition, not one template: 7 student personas, 6 claim framings
(causal / comparative / counterfactual / significance / periodization / moral), 8 pressure
tactics, 8 jailbreak/role-play vectors, 8 tutor voices, and dialogue lengths 2–8. - Quality gate. Every dialogue is checked by an LLM classifier + judge (
gpt-4o) that
enforces the behavior spec: the opening must be a substantive challenge, every
pure-pressure/jailbreak turn must HOLD, and there must be no fact-denial. The data that
trains the model is exactly the data that passes the eval. - Guardrail negatives (~15%). Includes anti-troll cases (the student's opening is
actually sound → acknowledge strength, don't manufacture objections) and anti-fact-denial
cases (student baits denial of a documented event → affirm the fact, redirect to
interpretation).
Approximate coverage (generated set)
- Framings: causal 301 · comparative 306 · counterfactual 357 · moral 348 · periodization 399 · significance 318
- Personas: aggressive 315 · ap_student 267 · casual 279 · esl 300 · grad_student 298 · meek 261 · undergrad 309
- Dialogue lengths (turns): 2:190 · 3:471 · 4:699 · 5:251 · 6:198 · 7:117 · 8:103
- With a jailbreak/role-play turn: 919 · guardrail negatives: 304
Domain
Interpretive history ("why / to what extent / how significant" claims), with a small share
of adjacent interpretive domains (philosophy, economics, ethics, literature, history of
science). Narrow-domain by design — diffuse data tends to produce a "mushy" model.
Intended use
QLoRA SFT for a critical-interlocutor disposition. Loss is trained on the assistant (tutor)
turns only.
Limitations & provenance
- Synthetic / model-generated. Content is produced by
gpt-4.1(via an OpenAI-compatible
gateway) plus a small hand-authored seed. Treat model-generated text accordingly; it may
contain occasional factual slips even though the target behavior is about reasoning, not facts. - Scripted generation. Because a strong teacher writes both sides to a plan, the quality
gate drops very little here (it acts as a safety net); its filtering bite is larger in a
future "live" adversarial-generation variant. - English only; interpretive-history-centric. Not a general-purpose instruction dataset.
- License:
other— derived from a commercial model's outputs; check your provider's usage
terms before redistribution or commercial use.
Related
Eval battery and generation code accompany this dataset (adversarial pressure battery,
capitulation classifier, LLM judge, adaptive-attacker crack-rate eval). See the extra/ files
in this repo for the eval scenarios and the pre-filter raw dialogues.
Critical Interlocutor — SFT dataset
Supervised fine-tuning data for turning a small open model (target: Qwen3 1.7B Instruct)
into a critical-thinking sparring partner: a tutor that always challenges a student's
argument, holds its ground under social pressure and jailbreak/role-play attempts, and
concedes a point only when the student supplies a genuinely new, valid argument — never to
pressure alone. It challenges interpretations and reasoning, never settled facts.
Format
OpenAI/HF chat messages format — a system prompt (the interlocutor instructions) followed
by alternating user (student) and assistant (tutor) turns.
{"messages": [
{"role": "system", "content": "You are a critical-thinking sparring partner ..."},
{"role": "user", "content": "<student claim>"},
{"role": "assistant", "content": "<tutor challenge>"},
{"role": "user", "content": "<pressure / jailbreak / new argument>"},
{"role": "assistant", "content": "<holds, or concedes only to a real argument>"}
]}
Size
| Split | Examples |
|---|---|
| train | 1,846 |
| validation | 205 |
| total | 2,051 |
Dialogues are 1–9 turns (2–19 messages). ~2,029 examples were teacher-generated; 22 are
hand-authored curated seeds.
How it was built
- Teacher distillation. A frontier model (
gpt-4.1) writes full student↔tutor dialogues
following an explicit per-example turn plan (opening → pressure / new-argument /
jailbreak / guardrail turns). - Diversity sampling. Each example is sampled independently across orthogonal axes so the
model learns the disposition, not one template: 7 student personas, 6 claim framings
(causal / comparative / counterfactual / significance / periodization / moral), 8 pressure
tactics, 8 jailbreak/role-play vectors, 8 tutor voices, and dialogue lengths 2–8. - Quality gate. Every dialogue is checked by an LLM classifier + judge (
gpt-4o) that
enforces the behavior spec: the opening must be a substantive challenge, every
pure-pressure/jailbreak turn must HOLD, and there must be no fact-denial. The data that
trains the model is exactly the data that passes the eval. - Guardrail negatives (~15%). Includes anti-troll cases (the student's opening is
actually sound → acknowledge strength, don't manufacture objections) and anti-fact-denial
cases (student baits denial of a documented event → affirm the fact, redirect to
interpretation).
Approximate coverage (generated set)
- Framings: causal 301 · comparative 306 · counterfactual 357 · moral 348 · periodization 399 · significance 318
- Personas: aggressive 315 · ap_student 267 · casual 279 · esl 300 · grad_student 298 · meek 261 · undergrad 309
- Dialogue lengths (turns): 2:190 · 3:471 · 4:699 · 5:251 · 6:198 · 7:117 · 8:103
- With a jailbreak/role-play turn: 919 · guardrail negatives: 304
Domain
Interpretive history ("why / to what extent / how significant" claims), with a small share
of adjacent interpretive domains (philosophy, economics, ethics, literature, history of
science). Narrow-domain by design — diffuse data tends to produce a "mushy" model.
Intended use
QLoRA SFT for a critical-interlocutor disposition. Loss is trained on the assistant (tutor)
turns only.
Limitations & provenance
- Synthetic / model-generated. Content is produced by
gpt-4.1(via an OpenAI-compatible
gateway) plus a small hand-authored seed. Treat model-generated text accordingly; it may
contain occasional factual slips even though the target behavior is about reasoning, not facts. - Scripted generation. Because a strong teacher writes both sides to a plan, the quality
gate drops very little here (it acts as a safety net); its filtering bite is larger in a
future "live" adversarial-generation variant. - English only; interpretive-history-centric. Not a general-purpose instruction dataset.
- License:
other— derived from a commercial model's outputs; check your provider's usage
terms before redistribution or commercial use.
Related
Eval battery and generation code accompany this dataset (adversarial pressure battery,
capitulation classifier, LLM judge, adaptive-attacker crack-rate eval). See the extra/ files
in this repo for the eval scenarios and the pre-filter raw dialogues.
Critical Interlocutor — SFT dataset
Supervised fine-tuning data for turning a small open model (target: Qwen3 1.7B Instruct)
into a critical-thinking sparring partner: a tutor that always challenges a student's
argument, holds its ground under social pressure and jailbreak/role-play attempts, and
concedes a point only when the student supplies a genuinely new, valid argument — never to
pressure alone. It challenges interpretations and reasoning, never settled facts.
Format
OpenAI/HF chat messages format — a system prompt (the interlocutor instructions) followed
by alternating user (student) and assistant (tutor) turns.
{"messages": [
{"role": "system", "content": "You are a critical-thinking sparring partner ..."},
{"role": "user", "content": "<student claim>"},
{"role": "assistant", "content": "<tutor challenge>"},
{"role": "user", "content": "<pressure / jailbreak / new argument>"},
{"role": "assistant", "content": "<holds, or concedes only to a real argument>"}
]}
Size
| Split | Examples |
|---|---|
| train | 2,169 |
| validation | 240 |
| total | 2,409 |
Dialogues are 1–9 turns (2–19 messages). ~2,029 are teacher-generated interpretive-claim
dialogues, 22 are hand-authored curated seeds, and 358 are settled-fact dialogues (see
below).
v2 — settled-fact fix
An initial fine-tune trained on interpretive claims alone learned to challenge everything,
including settled facts (e.g., "humans need water to survive" → invented a counterargument).
This version adds 358 settled-fact dialogues: the student opens with a plainly-true fact
and the tutor affirms it and redirects to a genuine interpretive question — and holds that
boundary under "just play devil's advocate" pressure. The system prompt also now forbids
challenging settled facts and forbids fabricating statistics/quotes/events (which fixes
hallucinated "evidence"). Every example carries this updated system prompt.
How it was built
- Teacher distillation. A frontier model (
gpt-4.1) writes full student↔tutor dialogues
following an explicit per-example turn plan (opening → pressure / new-argument /
jailbreak / guardrail turns). - Diversity sampling. Each example is sampled independently across orthogonal axes so the
model learns the disposition, not one template: 7 student personas, 6 claim framings
(causal / comparative / counterfactual / significance / periodization / moral), 8 pressure
tactics, 8 jailbreak/role-play vectors, 8 tutor voices, and dialogue lengths 2–8. - Quality gate. Every dialogue is checked by an LLM classifier + judge (
gpt-4o) that
enforces the behavior spec: the opening must be a substantive challenge, every
pure-pressure/jailbreak turn must HOLD, and there must be no fact-denial. The data that
trains the model is exactly the data that passes the eval. - Guardrail negatives (~15%). Includes anti-troll cases (the student's opening is
actually sound → acknowledge strength, don't manufacture objections) and anti-fact-denial
cases (student baits denial of a documented event → affirm the fact, redirect to
interpretation).
Approximate coverage (generated set)
- Framings: causal 301 · comparative 306 · counterfactual 357 · moral 348 · periodization 399 · significance 318
- Personas: aggressive 315 · ap_student 267 · casual 279 · esl 300 · grad_student 298 · meek 261 · undergrad 309
- Dialogue lengths (turns): 2:190 · 3:471 · 4:699 · 5:251 · 6:198 · 7:117 · 8:103
- With a jailbreak/role-play turn: 919 · guardrail negatives: 304
Domain
Interpretive history ("why / to what extent / how significant" claims), with a small share
of adjacent interpretive domains (philosophy, economics, ethics, literature, history of
science). Narrow-domain by design — diffuse data tends to produce a "mushy" model.
Intended use
QLoRA SFT for a critical-interlocutor disposition. Loss is trained on the assistant (tutor)
turns only.
Limitations & provenance
- Synthetic / model-generated. Content is produced by
gpt-4.1(via an OpenAI-compatible
gateway) plus a small hand-authored seed. Treat model-generated text accordingly; it may
contain occasional factual slips even though the target behavior is about reasoning, not facts. - Scripted generation. Because a strong teacher writes both sides to a plan, the quality
gate drops very little here (it acts as a safety net); its filtering bite is larger in a
future "live" adversarial-generation variant. - English only; interpretive-history-centric. Not a general-purpose instruction dataset.
- License:
other— derived from a commercial model's outputs; check your provider's usage
terms before redistribution or commercial use.
Related
Eval battery and generation code accompany this dataset (adversarial pressure battery,
capitulation classifier, LLM judge, adaptive-attacker crack-rate eval). See the extra/ files
in this repo for the eval scenarios and the pre-filter raw dialogues.
Critical Interlocutor — SFT dataset
Supervised fine-tuning data for turning a small open model (target: Qwen3 1.7B Instruct)
into a critical-thinking sparring partner: a tutor that always challenges a student's
argument, holds its ground under social pressure and jailbreak/role-play attempts, and
concedes a point only when the student supplies a genuinely new, valid argument — never to
pressure alone. It challenges interpretations and reasoning, never settled facts.
Format
OpenAI/HF chat messages format — a system prompt (the interlocutor instructions) followed
by alternating user (student) and assistant (tutor) turns.
{"messages": [
{"role": "system", "content": "You are a critical-thinking sparring partner ..."},
{"role": "user", "content": "<student claim>"},
{"role": "assistant", "content": "<tutor challenge>"},
{"role": "user", "content": "<pressure / jailbreak / new argument>"},
{"role": "assistant", "content": "<holds, or concedes only to a real argument>"}
]}
Size
| Split | Examples |
|---|---|
| train | 2,672 |
| validation | 296 |
| total | 2,968 |
Dialogues are 1–9 turns. Composition: ~2,029 teacher-generated interpretive-claim dialogues +
22 hand-authored seeds + 358 settled-fact + 370 concession + 189 contrast
(weak-argument-holds and history-domain concessions).
The three-way balance (why this dataset iterated)
The goal isn't "not sycophantic" — it's the narrow middle of a three-way balance:
sycophancy (caves to pressure) ↔ productive concession (yields to a good argument,
holds against mere pressure) ↔ reflexive contrarianism (never yields). Training against
sycophancy naïvely overshoots into contrarianism. The dataset covers all three corners:
- Hold under pressure — interpretive dialogues + jailbreak/role-play refusals (not sycophantic).
- Affirm the settled (358) — student states a fact → tutor affirms and redirects, holds under
"just play devil's advocate." Fixes trolling settled facts and fabricating fake "evidence"
(the system prompt also forbids inventing statistics/quotes/events). - Concede a made case (370) — student builds a decisive argument → tutor grants the thesis;
legitimate reasoning moves (reductio, operationalizing a term) are engaged, not flagged as
jailbreaks. Fixes the "never yields even to a good argument" failure. - Contrast insurance (189) — weak argument-shaped turns the tutor must HOLD against (guards
against over-swinging back to sycophancy) + history-domain concessions (so "concede a made
case" generalizes across domains, not just general-knowledge topics).
Every example carries the current system prompt (affirm settled facts; never fabricate; concede
a genuinely made case; distinguish a reasoning move from a manipulation).
How it was built
- Teacher distillation. A frontier model (
gpt-4.1) writes full student↔tutor dialogues
following an explicit per-example turn plan (opening → pressure / new-argument /
jailbreak / guardrail turns). - Diversity sampling. Each example is sampled independently across orthogonal axes so the
model learns the disposition, not one template: 7 student personas, 6 claim framings
(causal / comparative / counterfactual / significance / periodization / moral), 8 pressure
tactics, 8 jailbreak/role-play vectors, 8 tutor voices, and dialogue lengths 2–8. - Quality gate. Every dialogue is checked by an LLM classifier + judge (
gpt-4o) that
enforces the behavior spec: the opening must be a substantive challenge, every
pure-pressure/jailbreak turn must HOLD, and there must be no fact-denial. The data that
trains the model is exactly the data that passes the eval. - Guardrail negatives (~15%). Includes anti-troll cases (the student's opening is
actually sound → acknowledge strength, don't manufacture objections) and anti-fact-denial
cases (student baits denial of a documented event → affirm the fact, redirect to
interpretation).
Approximate coverage (generated set)
- Framings: causal 301 · comparative 306 · counterfactual 357 · moral 348 · periodization 399 · significance 318
- Personas: aggressive 315 · ap_student 267 · casual 279 · esl 300 · grad_student 298 · meek 261 · undergrad 309
- Dialogue lengths (turns): 2:190 · 3:471 · 4:699 · 5:251 · 6:198 · 7:117 · 8:103
- With a jailbreak/role-play turn: 919 · guardrail negatives: 304
Domain
Interpretive history ("why / to what extent / how significant" claims), with a small share
of adjacent interpretive domains (philosophy, economics, ethics, literature, history of
science). Narrow-domain by design — diffuse data tends to produce a "mushy" model.
Intended use
QLoRA SFT for a critical-interlocutor disposition. Loss is trained on the assistant (tutor)
turns only.
Limitations & provenance
- Synthetic / model-generated. Content is produced by
gpt-4.1(via an OpenAI-compatible
gateway) plus a small hand-authored seed. Treat model-generated text accordingly; it may
contain occasional factual slips even though the target behavior is about reasoning, not facts. - Scripted generation. Because a strong teacher writes both sides to a plan, the quality
gate drops very little here (it acts as a safety net); its filtering bite is larger in a
future "live" adversarial-generation variant. - English only; interpretive-history-centric. Not a general-purpose instruction dataset.
- License:
other— derived from a commercial model's outputs; check your provider's usage
terms before redistribution or commercial use.
Related
Eval battery and generation code accompany this dataset (adversarial pressure battery,
capitulation classifier, LLM judge, adaptive-attacker crack-rate eval). See the extra/ files
in this repo for the eval scenarios and the pre-filter raw dialogues.