Davd-b01/thinking-cap-tier-curricula-complete

Thinking Cap Tier Curricula — Complete Reasoning Alignment Suite (TCS v4)

[!IMPORTANT] Dataset Release v1.2 (Sept 2026) — Clean Delimiters & Zero-Padding Architecture: In v1.2, all 13,477 SFT samples and 3,187 SimPO preference pairs have undergone an automated token purge: - Zero `<|pad|>` batch residues: 100% eliminated across all files. - Zero reasoning leakage into final answers: Deliberation stays strictly inside <think>...</think>, and answers provide direct, non-repetitive conclusions. - Native ChatML tokens: Fully standardized on native atomic <think> and </think> tokens without legacy intermediate wrapper tags. - Clean sequence termination: Every sequence cleanly terminates with <|im_end|>.

This repository provides the complete, fully audited, and pedagogically ordered reasoning curriculum for post-training language models under the Thinking Cap Tiers (TCS v4 Cognitive Governance Standard).

Totaling ~36.3M tokens (22.8M SFT + 13.5M SimPO), this suite equips models with dynamic cognitive depth governance: the ability to seamlessly transition from immediate, direct answers (effort=off) to 4-phase formal deliberative proofs (effort=xhigh), reinforced by contrastive preference alignment without behavioral degradation.

1. Composition & Key Figures

• —SFT Curricula (`qwen_sft_curricula_v4.jsonl`): 13,477 curated traces (~22.82M tokens, 91.97 MB)
• —SimPO Preferences (`qwen_simpo_preference_v4.jsonl`): 3,187 contrastive pairs (~13.48M tokens, 50.88 MB)
• —Total Volume: ~36.30M tokens (142.85 MB)
• —Quality Audit: 100% Pass (0 corrupt JSONs, 0 legacy tags, 0 unclosed thoughts, 100% <|im_end|> termination).

2. The Thinking Cap Tiers (TCS v4 Cognitive Architecture)

Models trained with this curriculum obey an explicit system directive specifying the reasoning budget:

3. Delimiter & Tokenizer Contract

All traces are formatted in standard ChatML with native atomic reasoning tokens:

• —Introspection Open: <think>
• —Introspection Close: </think>
• —Turn Delimiters: <|im_start|> and <|im_end|>

SFT Data Format (qwen_sft_curricula_v4.jsonl):

Ordered identically to the RAW traces schema (prompt, think, answer, tier, domain):

{
  "prompt": "Find all real solutions to x^3 - 3x = 1.",
  "think": "[Phase 1: Framing]\nLet the polynomial be P(x) = x^3 - 3x - 1...\n[Phase 2: Dual Verification]...",
  "answer": "The real roots are x = 2*cos(20 deg), x = 2*cos(140 deg), and x = 2*cos(260 deg).",
  "tier": "high",
  "domain": "sft_math",
  "gold_solution": "x = 2*cos(20 deg), x = 2*cos(140 deg), and x = 2*cos(260 deg).",
  "seed_id": "SEED_MATH_FORMAL_00197_3f5d71e2",
  "complexity_score": 3.8,
  "text": "<|im_start|>system\nReasoning effort is set to high. Provide rigorous proof with verification.<|im_end|>\n<|im_start|>user\nFind all real solutions to x^3 - 3x = 1.<|im_end|>\n<|im_start|>assistant\n<think>\n[Phase 1: Framing]...\n</think>\nThe real roots are...<|im_end|>"
}

OFF-Mode Pure Brake:

In OFF mode, the output contains zero introspection tags:

<|im_start|>system
Reasoning effort is set to off. Answer directly and concisely.<|im_end|>
<|im_start|>user
Find 1 + 2 * 3 - 4 + 5.<|im_end|>
<|im_start|>assistant
8<|im_end|>

4. SimPO Length-Normalized Preference Alignment

Traditional DPO often suffers from length bias: models learn that longer completions receive higher cumulative reward, leading to catastrophic verbosity on trivial queries.

This dataset implements SimPO (Simple Preference Optimization, Meng et al., NeurIPS 2024):

$$ \mathcal{L}{\text{SimPO}}(\theta) = -\mathbb{E}{(x, yw, yl)} \left[ \log \sigma \left( \frac{\beta}{|yw|} \log \pi\theta(yw \mid x) - \frac{\beta}{|yl|} \log \pi\theta(yl \mid x) - \gamma \right) \right] $$

SimPO Data Format (qwen_simpo_preference_v4.jsonl):

Clean separation of prompt, chosen deliberation, chosen answer, and contrastive rejected elements:

{
  "prompt": "What is dry ice?",
  "chosen_think": "The user wants a definition of 'dry ice.' I need to provide its identity, physical properties, and common applications.",
  "chosen_answer": "Dry ice is the solid form of carbon dioxide (CO2), which sublimates directly from solid to gas at -78.5°C.",
  "rejected_think": "1. Deconstruct the Request: Multi-phase analysis on basic factual definition...",
  "rejected_answer": "Dry ice is the solid crystalline form of carbon dioxide...",
  "type": "anti_overthinking_conciseness",
  "domain": "sft_dialogue",
  "system_prompt": "Reasoning effort is set to low. Think briefly, then answer.",
  "id": "SIMPO_ANTI_OVERTHINK_00412"
}

Contrastive Pair Typology:

1. 1.Anti-Overthinking Pairs (2,328 pairs):
2. 2.Evaluated on low-complexity prompts under effort=low or effort=off.
3. 3.chosen ($y_w$): Concise, accurate response (~20–60 words).
4. 4.rejected ($y_l$): Verbose multi-phase over-deliberation on a simple greeting or arithmetic task (~400 words).
5. 5.Result: The policy learns that unneeded introspection on trivial tasks incurs a severe reward penalty.
6. 6.Rigorous Verification Pairs (859 pairs):
7. 7.Evaluated on Olympiad math and complex code under effort=high.
8. 8.chosen ($y_w$): Thorough dual-check proof that catches subtle edge cases.
9. 9.rejected ($y_l$): Superficial, single-pass deduction that misses boundary conditions.

5. Domain Distribution

The SFT curriculum spans 7 rigorously audited domains:

• —Math Formal (35.5%): Olympiad, AMC12, Putnam, Algebra, Geometry, Calculus with SymPy verification.
• —Science Logic & Data (33.5%): Physics, Chemistry, Biology, Causal Inference, Data synthesis.
• —Executable Code (20.6%): Python, Rust, SQL, algorithms, all with AST compilation pass.
• —Strict Instruction / IFEval (3.9%): Negative constraints, word limits, exact formatting.
• —Dialogue & Philosophy (3.2%): Nuanced humanities discussions without artificial refusal tokens.
• —Agentic Tools (1.8%): Function calling, schema validation, tool execution reasoning.
• —Multilingual Native (1.5%): High-depth reasoning natively composed in Spanish, English, and Chinese.

6. How to Use

Loading SFT and SimPO Splits:

from datasets import load_dataset

# Load SFT Curriculum (13,477 curated traces)
sft_ds = load_dataset("Davd-b01/thinking-cap-tier-curricula-complete", "sft")
print("SFT samples:", len(sft_ds["train"]))

# Load SimPO Preference Pairs (3,187 contrastive pairs)
simpo_ds = load_dataset("Davd-b01/thinking-cap-tier-curricula-complete", "simpo")
print("SimPO pairs:", len(simpo_ds["train"]))

🎯 What is Sought in Each Thinking Tier? (Cognitive Architecture & Objectives)

The Thinking Cap Tiers (TCS v4 Cognitive Governance Standard) enforces explicit behavioral contracts across 5 tiers:

1. Tier OFF (effort=off): The Immediate Brake

• —Cognitive Objective: Zero introspection, zero hallucinated meta-cognition, strict conciseness.
• —The Problem Solved: Traditional reasoning models frequently over-deliberate on simple arithmetic ($1+1$) or polite greetings ("Hello").
• —Enforced Behavior: Emits the exact answer directly (≤ 50 words), terminating dryly with <|im_end|>.

2. Tier LOW (effort=low): Agile Unilinear Deduction

• —Cognitive Objective: Fast, single-pass linear derivation with instant self-check.
• —Enforced Behavior: Generates a brief thought trace (~150–250 words) without alternative branching or verbose preambles.

3. Tier MID (effort=mid): Structured Pedagogical Exposition

• —Cognitive Objective: Clear step-by-step pedagogical clarity, transparent intermediate states.
• —Enforced Behavior: Generates a structured thought trace (~300–500 words) detailing the rationale behind algebraic steps and invariants.

4. Tier HIGH (effort=high): Formal Dual-Branch Proof

• —Cognitive Objective: Rigorous formal proof combined with a mandatory secondary verification branch.
• —Enforced Behavior: Generates a comprehensive trace (~600–900 words) with explicit checks on boundary conditions and edge cases.

5. Tier XHIGH (effort=xhigh): 4-Phase Deep Deliberation

• —Cognitive Objective: Deep philosophical and mathematical deliberation structured across 4 explicit cognitive phases:
• —Phase 1 (Framing & Invariant Discovery)
• —Phase 2 (Hypothesis & Counter-Example Search)
• —Phase 3 (Deep Formal Proof / Synthesis)
• —Phase 4 (Stress-Testing & Boundary Synthesis)

🏛️ Trace Generators, Attribution & Upstream Acknowledgments

We gratefully acknowledge the foundational teacher models and upstream datasets that generated these traces:

1. 1.`r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation` (r0b0tlab): Anchor seed problems and teacher reasoning generated by Qwen 3.8 Max, GLM 5.2, and Moonshot Kimi k3.
2. 2.OpenThoughts Dataset Collection (`open-thoughts/OpenThoughts-114k`): Formal mathematical problem distributions and deliberative structures.
3. 3.OpenMLE-SFT & Bespoke-Stratos Collections: Real-world execution-grounded software engineering traces.
4. 4.LIMA Research (Zhou et al., NeurIPS 2023): Surface alignment hypothesis.

🔗 Related Thinking Cap Datasets

• —⚡ [Thinking Cap LIMA Hyper-Dense](https://huggingface.co/datasets/Davd-b01/thinking-cap-tier-lima-dense): Top 5.5k SFT + 2k SimPO pairs for budget-friendly alignment.
• —📦 [Thinking Cap Tier Raw Traces](https://huggingface.co/datasets/Davd-b01/thinking-cap-tier-raw-traces): 38,158 raw candidate generation outputs across all tiers.

License & Citation

This dataset is released under the Apache 2.0 License.

@dataset{thinking_cap_tier_curricula_complete_2026,
  author = {Davd-b01},
  title = {Thinking Cap Tier Curricula: Complete Reasoning Alignment Suite (TCS v4)},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/Davd-b01/thinking-cap-tier-curricula-complete}
}