保存先:
| 第一著者: | |
|---|---|
| フォーマット: | Recurso digital |
| 言語: | 英語 |
| 出版事項: |
Zenodo
2026
|
| 主題: | |
| オンライン･アクセス: | https://doi.org/10.5281/zenodo.18967869 |
| タグ: |
タグ追加
タグなし, このレコードへの初めてのタグを付けませんか!
|
目次:
- <p>This technical report presents <strong>Natural-Synthesis-8B</strong>, an experimental fine-tune of Llama-3-8B trained on a synthetic dataset of 68 examples designed to install a biologically-inspired reasoning paradigm rather than domain-specific knowledge.</p> <p>The central hypothesis is that reasoning <em>process</em> and reasoning <em>content</em> are separable learning targets. Standard fine-tuning teaches models what to think. This work teaches a model <em>how</em> to think — specifically, through a five-stage cognitive growth cycle (Seed → Root Exploration → Principled Pruning → Canopy Formation → Homeostatic Review) governed by five evaluative nutrients (Coherence, Parsimony, Explanatory Power, Fecundity, Evidential Grounding).</p> <p>The result is a model that demonstrates consistent cross-domain structural reasoning, emergent systems thinking, and selective metacognitive activation at 8B parameters — capabilities that do not appear reliably in the base model. Custom evaluations show an 18% gain in cognitive flexibility over the base model baseline.</p> <p>This report documents the paradigm, training methodology, benchmark comparisons, qualitative behavioral evidence, and a frank analysis of failure modes — including the coherence-without-truth problem inherent to any coherence-optimized reasoning system.</p> <p>The model is available at: https://huggingface.co/JPQ24/llama-3-8b-Natural-synthesis-Lora-Merge</p> <p>The dataset is aviable at: https://huggingface.co/datasets/JPQ24/Natural_synthesis</p>