reaperdoesntknow/DualMind

1---2license: apache-2.03library_name: transformers4pipeline_tag: text-generation5tags:6- qwen37- sft8- trl9- dual-mind10- reasoning11- convergent-intelligence12- explore-examine-response13- convergentintel14- edge15- distillation16- knowledge-distillation17datasets:18- zai-org/LongWriter-6k19base_model:20- reaperdoesntknow/DiStil-Qwen3-1.7B-uncensored21---22 23# DualMind24 25**Single Architecture, Dual Cognition — The Multi-Model Collision Array on Shared Weights**26 27*Convergent Intelligence LLC: Research Division*28 29---30 31## What This Is32 33DualMind 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:34 35- **`<explore>`** — Unconstrained reasoning. Derivation, speculation, working through the problem freely.36- **`<examine>`** — Adversarial self-response. The model reads its own explore output and critiques it. Error detection, verification, refinement.37- **`<response>`** — Clean synthesis. The final answer distilled from the internal dialogue.38 39This 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](https://huggingface.co/reaperdoesntknow)) is recreated through role-conditioned generation on shared weights.40 41## Architecture42 43| Parameter | Value |44|-----------|-------|45| Architecture | Qwen3ForCausalLM |46| Parameters | ~2.03B (1.7B effective) |47| Hidden Size | 2048 |48| Layers | 28 |49| Attention Heads | 16 (Q) / 8 (KV) — GQA |50| Context Length | 40,960 tokens |51| Precision | BF16 (trained on H100) |52 53## Training54 55**Base model:** [Disctil-Qwen3-1.7B](https://huggingface.co/reaperdoesntknow/Disctil-Qwen3-1.7B) (DISC-refined uncensored Qwen3)56 57**Dataset:** [KK04/LogicInference_OA](https://huggingface.co/datasets/KK04/LogicInference_OA) — Logical inference problems transformed into the DualMind cognitive loop format.58 59**Training format:** Each CoT solution is restructured into the DualMind format:60- Derivation sentences → `<explore>` block (reasoning phase)61- Verification/checking sentences → `<examine>` block (self-critique phase)62- Final answer → `<response>` block (synthesis)63 64Sentence-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.65 66**Hardware:** Colab H100, BF16 precision. 512 steps, lr 5e-6, SFT via TRL.67 68**Next iteration:** Currently training on [Crownelius/Opus-4.6-Reasoning-3300x](https://huggingface.co/datasets/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.69 70## Usage71 72```python73from transformers import AutoModelForCausalLM, AutoTokenizer74 75model = AutoModelForCausalLM.from_pretrained(76    "reaperdoesntknow/DualMind",77    torch_dtype="auto",78    device_map="auto"79)80tokenizer = AutoTokenizer.from_pretrained("reaperdoesntknow/DualMind")81 82# Start the explore block — the model completes the full loop83prompt = (84    "##USER:\n"85    "Prove that the sum of two even numbers is always even.\n\n"86    "<explore>\n"87)88 89inputs = tokenizer(prompt, return_tensors="pt").to(model.device)90output = model.generate(91    **inputs,92    max_new_tokens=1024,93    do_sample=True,94    top_p=0.9,95    temperature=0.6,96    repetition_penalty=1.15,97)98result = tokenizer.decode(output[0], skip_special_tokens=True)99print(result)100```101 102### Expected Output Structure103 104```105<explore>106[The model works through the proof freely — definitions, algebraic manipulation, etc.]107</explore>108 109<examine>110[The model critiques its own derivation — checks for gaps, verifies steps, catches errors]111</examine>112 113<response>114[Clean final answer synthesized from the internal dialogue]115</response>116```117 118## Why Dual Modality119 120Standard 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:121 1221. **Explore** is free to make mistakes, speculate, and try approaches that might not work1232. **Examine** reads the explore output adversarially — it's looking for errors, not confirming correctness1243. **Response** has the benefit of both perspectives125 126This 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.127 128## Distillation Chain129 130```131Qwen3-1.7B (base)132  → DiStil-Qwen3-1.7B-uncensored (uncensored SFT)133    → Disctil-Qwen3-1.7B (DISC refinement)134      → DualMind (DualMind SFT on Opus 4.6 reasoning data) ← you are here135```136 137 138## Mathematical Foundations: Discrepancy Calculus (DISC)139 140DualMind'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.141 142The discrepancy operator:143 144$$Df(x) = \lim_{\varepsilon \downarrow 0} \frac{1}{\varepsilon} \int_x^{x+\varepsilon} \frac{|f(t) - f(x)|}{|t - x|}\, dt$$145 146quantifies 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.147 148Full theory: *"On the Formal Analysis of Discrepancy Calculus"* (CIx, 2026; Convergent Intelligence LLC: Research Division).149 150## Related Models151 152| Model | Description | Downloads |153|-------|-------------|-----------|154| [TopologicalQwen](https://huggingface.co/reaperdoesntknow/TopologicalQwen) | TKD + DualMind on physics CoT | 622 |155| [Disctil-Qwen3-1.7B](https://huggingface.co/reaperdoesntknow/Disctil-Qwen3-1.7B) | Parent model (DISC-refined) | 286 |156| [Qwen3-1.7B-Thinking-Distil](https://huggingface.co/reaperdoesntknow/Qwen3-1.7B-Thinking-Distil) | TKD with Thinking teacher | 687 |157 158**[DualMind Collection](https://huggingface.co/collections/reaperdoesntknow/dualmind)** — Dual-cognition model series159 160**[DistilQwen Collection](https://huggingface.co/collections/reaperdoesntknow/distilqwen-69bf40ec669117e3f069ef1c)** — Full proof-weighted distillation series161 162Full methodology: [Structure Over Scale (DOI: 10.57967/hf/8165)](https://doi.org/10.57967/hf/8165)163 164## Citation165 166```bibtex167@misc{cix2026dualmind,168  title={DualMind: Dual-Mental-Modality Reasoning via Role-Conditioned Self-Critique},169  author={Convergent Intelligence},170  year={2026},171  publisher={HuggingFace},172  url={https://huggingface.co/reaperdoesntknow/DualMind},173  note={Convergent Intelligence LLC: Research Division}174}175```176 177---178 179*Convergent Intelligence LLC: Research Division*180*"Where classical analysis fails to see, we begin."*181<!-- cix-keeper-ts:2026-09-23T13:15:28Z -->182<!-- card-refresh: 2026-03-30 -->183 184---185 186## Convergent Intelligence Portfolio187 188*Part of the [DualMind Series](https://huggingface.co/collections/reaperdoesntknow/dualmind-69c93f888c6e79ecc69cf41e) by [Convergent Intelligence LLC: Research Division](https://huggingface.co/reaperdoesntknow)*189 190### DualMind Family191 192| Model | Format | Description |193|-------|--------|-------------|194| [DualMind](https://huggingface.co/reaperdoesntknow/DualMind) | BF16 | LogicInference-trained. Explore→Examine→Response loop. |195| [DualMinded-Qwen3-1.7B](https://huggingface.co/reaperdoesntknow/DualMinded-Qwen3-1.7B) | BF16 | Opus 4.6 reasoning traces. Higher quality splits. |196| [Dualmind-Qwen-1.7B-Thinking](https://huggingface.co/reaperdoesntknow/Dualmind-Qwen-1.7B-Thinking) | BF16 | Thinking-teacher variant with extended deliberation. |197| [DualMind-GGUF](https://huggingface.co/reaperdoesntknow/DualMind-GGUF) | GGUF | Quantized LogicInference variant. CPU/6GB GPU. |198| [DualMinded-Qwen3-1.7B-GGUF](https://huggingface.co/reaperdoesntknow/DualMinded-Qwen3-1.7B-GGUF) | GGUF | Quantized Opus variant. Ollama ready. |199 200### Papers201 202| Paper | DOI |203|-------|-----|204| [Structure Over Scale](https://huggingface.co/reaperdoesntknow/Structure-Over-Scale) | 10.57967/hf/8165 |205| [Three Teachers to Dual Cognition](https://huggingface.co/reaperdoesntknow/DualMind_Methodolgy) | 10.57967/hf/8184 |206| [Discrepancy Calculus](https://huggingface.co/reaperdoesntknow/Discrepancy_Calculus) | 10.57967/hf/8194 |207 208---209 210*Last updated: 2026-03-31 by Convergent Intelligence LLC: Research Division*211