README.md

metadata

license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
  - qwen3
  - sft
  - trl
  - dual-mind
  - reasoning
  - convergent-intelligence
  - explore-examine-response
datasets:
  - zai-org/LongWriter-6k
base_model:
  - reaperdoesntknow/DiStil-Qwen3-1.7B-uncensored

DualMind

Single Architecture, Dual Cognition — The Multi-Model Collision Array on Shared Weights

Convergent Intelligence LLC: Research Division

---

What This Is

DualMind is a 1.7B parameter model that implements dual-mental-modality reasoning — a single model with two internal voices sharing the same weights, differentiated only by role tokens:

• <explore> — Unconstrained reasoning. Derivation, speculation, working through the problem freely.
• <examine> — Adversarial self-response. The model reads its own explore output and critiques it. Error detection, verification, refinement.
• <response> — Clean synthesis. The final answer distilled from the internal dialogue.

This is the multi-model collision array collapsed into a single architecture. The dialectical structure that produces novel insights from architectural diversity (demonstrated in our five-architecture collision experiments) is recreated through role-conditioned generation on shared weights.

Architecture

Parameter	Value
Architecture	Qwen3ForCausalLM
Parameters	~2.03B (1.7B effective)
Hidden Size	2048
Layers	28
Attention Heads	16 (Q) / 8 (KV) — GQA
Context Length	40,960 tokens
Precision	BF16 (trained on H100)

Training

Base model: Disctil-Qwen3-1.7B (DISC-refined uncensored Qwen3)

Dataset: KK04/LogicInference_OA — Logical inference problems transformed into the DualMind cognitive loop format.

Training format: Each CoT solution is restructured into the DualMind format:

• Derivation sentences → <explore> block (reasoning phase)
• Verification/checking sentences → <examine> block (self-critique phase)
• Final answer → <response> block (synthesis)

Sentence-level splitting uses trigger detection (check, verify, however, but wait, etc.) to find the natural transition from reasoning to verification, with 70/30 positional fallback.

Hardware: Colab H100, BF16 precision. 512 steps, lr 5e-6, SFT via TRL.

Next iteration: Currently training on Crownelius/Opus-4.6-Reasoning-3300x — 2,160 Claude Opus 4.6 reasoning samples with pre-separated thinking/solution columns, eliminating the need for heuristic splitting.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "reaperdoesntknow/DualMind",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("reaperdoesntknow/DualMind")

# Start the explore block — the model completes the full loop
prompt = (
    "##USER:\n"
    "Prove that the sum of two even numbers is always even.\n\n"
    "<explore>\n"
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(
    **inputs,
    max_new_tokens=1024,
    do_sample=True,
    top_p=0.9,
    temperature=0.6,
    repetition_penalty=1.15,
)
result = tokenizer.decode(output[0], skip_special_tokens=True)
print(result)

Expected Output Structure

<explore>
[The model works through the proof freely — definitions, algebraic manipulation, etc.]
</explore>

<examine>
[The model critiques its own derivation — checks for gaps, verifies steps, catches errors]
</examine>

<response>
[Clean final answer synthesized from the internal dialogue]
</response>

Why Dual Modality

Standard CoT prompting produces a single stream of reasoning. The model has one shot to get it right. DualMind gives the model a structural mechanism for self-correction:

1. Explore is free to make mistakes, speculate, and try approaches that might not work
2. Examine reads the explore output adversarially — it's looking for errors, not confirming correctness
3. Response has the benefit of both perspectives

This mirrors what happens in multi-model collision arrays where different architectures produce genuinely different failure modes, and the collision between them surfaces structure that neither achieves alone. DualMind recreates this dynamic within a single set of weights through role conditioning.

Distillation Chain

Qwen3-1.7B (base)
  → DiStil-Qwen3-1.7B-uncensored (uncensored SFT)
    → Disctil-Qwen3-1.7B (DISC refinement)
      → DualMind (DualMind SFT on Opus 4.6 reasoning data) ← you are here

Mathematical Foundations: Discrepancy Calculus (DISC)

DualMind's dual-cognition architecture connects to Discrepancy Calculus through Continuous Thought Dynamics (Ch. 19 of the DISC monograph) — which models inference as a discrepancy-guided PDE where the explore→examine→respond cycle corresponds to a controlled trajectory through cognitive phase space.

The discrepancy operator:

$$Df(x)=\underset{\epsilon \downarrow 0}{\lim }\frac{1}{\epsilon }{\int }_x^{x+\epsilon }\frac{\mathrm{\mid }f(t)-f(x)\mathrm{\mid }}{\mathrm{\mid }t-x\mathrm{\mid }}dtDf(x)\; =\; \backslash lim\_\{\backslash varepsilon\; \backslash downarrow\; 0\}\; \backslash frac\{1\}\{\backslash varepsilon\}\; \backslash int\_x^\{x+\backslash varepsilon\}\; \backslash frac\{|f(t)\; -\; f(x)|\}\{|t\; -\; x|\}\backslash ,\; dt$$

quantifies the mismatch between what the model generates (integration) and what it should generate (differentiation). The <explore> phase increases discrepancy energy freely; <examine> applies the Adaptive Discrepancy Derivative (ADD, Ch. 14) to detect drift; <response> minimizes residual discrepancy into a clean output. The three phases implement the BV decomposition operationally: smooth reasoning, jump corrections at error boundaries, and Cantor-type refinement of subtle drift.

Full theory: "On the Formal Analysis of Discrepancy Calculus" (Colca, 2026; Convergent Intelligence LLC: Research Division).

Related Models

Model	Description	Downloads
TopologicalQwen	TKD + DualMind on physics CoT	622
Disctil-Qwen3-1.7B	Parent model (DISC-refined)	286
Qwen3-1.7B-Thinking-Distil	TKD with Thinking teacher	687

DualMind Collection — Dual-cognition model series

DistilQwen Collection — Full proof-weighted distillation series

Full methodology: Structure Over Scale (DOI: 10.57967/hf/8165)

Citation

@misc{colca2026dualmind,
  title={DualMind: Dual-Mental-Modality Reasoning via Role-Conditioned Self-Critique},
  author={Colca, Roy S.},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/reaperdoesntknow/DualMind},
  note={Convergent Intelligence LLC: Research Division}
}

---

Convergent Intelligence LLC: Research Division "Where classical analysis fails to see, we begin."