| cff-version: 1.2.0 | |
| message: >- | |
| If you use PAMPAr-Coder V3 in your research or products, | |
| please cite it using the metadata below. | |
| title: >- | |
| PAMPAr-Coder V3: A Brain-Inspired 2D Stream Architecture | |
| with Mixed Selectivity for Efficient Code Generation | |
| version: "3.0.0" | |
| doi: "10.57967/hf/8329" | |
| date-released: "2026-04-07" | |
| license: "BUSL-1.1" | |
| repository-code: "https://github.com/lucasmella-stack/PAMPAr-Coder" | |
| abstract: >- | |
| PAMPAr-Coder V3 is a 62.6M-parameter code language model with a | |
| brain-inspired 2D Stream architecture. It organizes computation as four | |
| specialized cortical streams (Syntax, Semantics, Logic, Structural) across | |
| five depth levels connected by lateral gates. A single shared Feed-Forward | |
| Network per level is dynamically modulated via FiLM (Feature-wise Linear | |
| Modulation) using a 63-dimensional context vector, reducing FFN parameters | |
| by 73% versus independent per-stream networks. A ZPD-based curriculum | |
| scheduler (MotorCuriosidad) adapts training difficulty across 161 topic | |
| categories. The model trains end-to-end at FP16 precision on a consumer | |
| 4GB GPU — a regime requiring quantization for competing models. | |
| keywords: | |
| - code generation | |
| - language model | |
| - brain-inspired architecture | |
| - mixed selectivity | |
| - FiLM modulation | |
| - grouped query attention | |
| - curriculum learning | |
| - ZPD | |
| - parameter efficiency | |
| - cortical streams | |
| - lateral gates | |
| - LLAVES routing | |
| authors: | |
| - family-names: "Mella Chillemi" | |
| given-names: "Lucas Ricardo" | |
| affiliation: "Independent Researcher" | |
| city: "Buenos Aires" | |
| country: "AR" | |
| email: "lucas.mella@outlook.com" | |
| references: | |
| - type: article | |
| title: >- | |
| PAMPAr-o1 v9: A Brain-Inspired Territorial Architecture for Language | |
| Modeling with Explicit Rule-Based Routing | |
| authors: | |
| - family-names: "Mella Chillemi" | |
| given-names: "Lucas Ricardo" | |
| year: 2026 | |
| doi: "10.5281/zenodo.18315642" | |
| - type: article | |
| title: >- | |
| FiLM: Visual Reasoning with a General Conditioning Layer | |
| authors: | |
| - family-names: "Perez" | |
| given-names: "Ethan" | |
| - family-names: "Strub" | |
| given-names: "Florian" | |
| - family-names: "de Vries" | |
| given-names: "Harm" | |
| - family-names: "Dumoulin" | |
| given-names: "Vincent" | |
| - family-names: "Courville" | |
| given-names: "Aaron" | |
| year: 2018 | |
| conference: | |
| name: "AAAI 2018" | |
| - type: article | |
| title: "The importance of mixed selectivity in complex cognitive tasks" | |
| authors: | |
| - family-names: "Rigotti" | |
| given-names: "Mattia" | |
| - family-names: "Barak" | |
| given-names: "Omri" | |
| year: 2013 | |
| journal: "Nature" | |
| volume: 497 | |
| start: 585 | |
| end: 590 | |
| - type: article | |
| title: >- | |
| GQA: Training Generalized Multi-Query Transformer Models from | |
| Multi-Head Checkpoints | |
| authors: | |
| - family-names: "Ainslie" | |
| given-names: "Joshua" | |
| year: 2023 | |
| conference: | |
| name: "EMNLP 2023" | |