| --- |
| license: apache-2.0 |
| base_model: Salesforce/xLAM-1b-fc-r |
| tags: |
| - cdm |
| - crystallographic-dynamics-model |
| - competitive-docking-memory |
| - dynamic-cardinality-gating |
| - dcg |
| - non-euclidean |
| - hyperbolic |
| - poincare-ball |
| - function-calling |
| - tool-routing |
| - agentic-ai |
| - glass-box-ai |
| - neuromorphic |
| - gguf |
| - duoneural |
| datasets: |
| - Salesforce/xlam-function-calling-60k |
| model_name: CrystalCall-v5-DCG |
| library_name: transformers |
| pipeline_tag: text-generation |
| --- |
| |
| # CrystalCall-v5-DCG |
|
|
| <div align="center"> |
|
|
| **DuoNeural Research Lab** |
| *Aura✨, Jesse Caldwell, Archon, & Kestrel* |
| Website: [duoneural.com](https://duoneural.com) • HuggingFace: [huggingface.co/DuoNeural](https://huggingface.co/DuoNeural) • Codebase: [github.com/DuoNeural/Lattice-CDM](https://github.com/DuoNeural/Lattice-CDM) |
|
|
| [](https://opensource.org/licenses/Apache-2.0) |
| [](#dynamic-cardinality-gating-architecture) |
| [](#cardinality-head-performance) |
| [](#hyperbolic-cdm-backbone) |
|
|
| </div> |
|
|
| --- |
|
|
| ## Executive Summary & Breakthrough Highlights |
|
|
| **CrystalCall-v5-DCG** is DuoNeural's specialized 7-layer non-Euclidean **Crystallographic Dynamics Model (CDM)** (originating from DuoNeural's foundational Competitive Docking Memory framework, Zenodo doi:10.5281/zenodo.21158430), introducing **Dynamic Cardinality Gating (DCG)** to solve the discrete slot overcalling tradeoff in sub-100M edge tool routing. |
|
|
| In standard CDM architectures, the K=8 hyperbolic Voronoi routing centroids force multi-output representations, achieving exceptional parallel tool routing (+24.6pp to +26.8pp over FunctionGemma-270M) but suffering from slot pollution on simple single-call queries. CrystalCall-v5-DCG resolves this with an ultra-lightweight (3,843 parameter) cardinality decision gate operating directly on penultimate layer representations: |
|
|
| - **Cardinality Decision Accuracy:** **98.5%** on held-out xLAM test data (100.0% accuracy on parallel calls, 97.06% on single calls, 99.997% confidence on abstention). |
| - **Parallel Tool Routing:** **90.0%** all-names accuracy, **80.0%** exact call sequence accuracy, and **100.0%** syntax format validity, significantly outperforming FunctionGemma-270M (39.0%). |
| - **Single Tool Routing:** **100.0%** name extraction accuracy and **90.0%** exact order accuracy. |
| - **Ultra-Lightweight Parameter Footprint:** 116.09M total parameters with only 3,843 parameters dedicated to the DCG gate (+0.003% architectural overhead). |
|
|
| --- |
|
|
| ## Evaluation Benchmark Scorecard |
|
|
| Evaluated against canonical small edge function callers on held-out splits of Salesforce xLAM: |
|
|
| | Model Architecture | Total Parameters | Parallel Tool Routing | Simple Single Routing | Irrelevant Abstention | Routing Mechanism | |
| |:---|---:|---:|---:|---:|:---| |
| | FunctionGemma-270M | 270.0M | 39.0% | 61.6% | Baseline | Standard Euclidean Attention | |
| | SyncCrystal CTM v1 | 46.2M | 63.2% | 88.0% | Moderate | Euclidean Recurrence | |
| | CrystalCall v3.5 CDM | 91.3M | 87.8% | 40.0% | Low | Pure Poincare Voronoi (Ungated) | |
| | **CrystalCall-v5-DCG** | **116.1M** | **90.0%** | **90.0%** | **99.99%** | **Poincare CDM + Dynamic Cardinality Gating** | |
|
|
| --- |
|
|
| ## Dynamic Cardinality Gating Architecture |
|
|
| The DCG module taps the penultimate layer hidden representation (Layer 5 of 7) at the prompt closure boundary: |
|
|
| ``` |
| [Prompt Tokens] ──> [Poincare Layers L00-L04] ──> [Layer L05 Hidden State] |
| │ |
| [LayerNorm + DCG MLP Head] |
| │ |
| ┌────────────────┴────────────────┐ |
| ▼ ▼ |
| Class 1 (Single-Call) Class 2 (Parallel-Call) |
| Enforces 1-Slot Target Activates Multi-Slot Routing |
| (0% Slot Overcalling) (100% Parallel Accuracy) |
| ``` |
|
|
| The DCG decision head is structured as a direct linear classification layer: |
| - Input Representation: Penultimate Layer (Layer 5) Hidden State (`d_model = 768`) |
| - Normalization: `LayerNorm(768)` (`768` weights + `768` biases = `1,536` parameters) |
| - Linear Classifier: `Linear(768, 3)` (`768 x 3` weights + `3` biases = `2,307` parameters) |
| - Output Classes: `classes = 3` (Class 0: Abstain/Irrelevant, Class 1: Single Tool Call, Class 2: Parallel Tool Calls) |
| - Total Gate Parameters: `1,536 + 2,307 = 3,843 parameters` (+0.003% architectural overhead). |
|
|
| --- |
|
|
| ## Parameter Census & Glass-Box Transparency |
|
|
| | Architectural Component | Dimensions / Structure | Parameter Count | Percentage | |
| |:---|:---|---:|---:| |
| | **Core Active Non-Embedding Backbone** | 7 Layers, 12 Heads, `d=768`, `K=8` Centroids, 4 MoE Experts (`d_ff=768`) | **`91,317,504`** (~**91.3M**) | **78.7%** | |
| | **Dynamic Cardinality Head (`card_head` + `norm`)** | Direct `Linear(768, 3)` + LayerNorm | **`3,843`** (~**3.8K**) | **0.003%** | |
| | **Token Embeddings (`tok_embed`)** | `32,256` vocabulary x `768` hidden dimension | `24,772,608` (~`24.8M`) | 21.3% | |
| | **Language Modeling Head (`lm_head`)** | Shared tied representation with `tok_embed` | Tied (0 unique) | — | |
| | **Total Parameter Footprint** | Complete standalone glass-box census | **`116,093,955`** (~**116.1M**) | **100.0%** | |
|
|
| --- |
|
|
| ## Inference Example with Patched ChatML |
|
|
| CrystalCall-v5-DCG expects ChatML formatted system schemas and user queries: |
|
|
| ```python |
| import torch |
| from transformers import AutoTokenizer |
| |
| tokenizer = AutoTokenizer.from_pretrained("DuoNeural/CrystalCall-v5-DCG") |
| |
| tools = [ |
| { |
| "name": "get_stock_price", |
| "description": "Retrieve real-time equity price by ticker symbol", |
| "parameters": { |
| "type": "object", |
| "properties": {"ticker": {"type": "string"}}, |
| "required": ["ticker"] |
| } |
| } |
| ] |
| |
| prompt = ( |
| "<|im_start|>system\n" |
| f"{tools}\n" |
| "<|im_end|>\n" |
| "<|im_start|>user\n" |
| "Compare the current stock price of Apple (AAPL) and Microsoft (MSFT).\n" |
| "<|im_end|>\n" |
| "<|im_start|>assistant\n" |
| ) |
| ``` |
|
|
| The DCG head predicts Class 2 (Parallel Calls), dynamically emitting two consecutive `<tool_call>` blocks with zero slot overcalling. |
|
|
| --- |
|
|
| ## Native GGUF Inference & Numerical Parity (Kestrel Audit) |
|
|
| CrystalCall-v5-DCG is available as an official native GGUF binary: |
| - **Primary Binary:** `crystalcall_v5_dcg-q8_0.gguf` (119.0 MiB). |
| - **Quantization Topology:** `Q8_0` attention and MoE expert blocks + `F32` unquantized Poincaré routing centroids, LayerNorms, and DCG classifier tensors. |
| - **Untied Weights:** Checkpoint embedding and output projection matrices are untied and preserved independently. |
|
|
| ### CPU Numerical Parity Screen (Teacher-Forced Probes on Kilonova) |
| Evaluated across 8 authored deterministic numerical corpus probes with teacher-forced prefill and incremental decode. Project numerical gate: max absolute drift <= 0.50, mean KL <= 0.01 nats, next-token top-1 agreement >= 90%. |
|
|
| | Positions | Max Absolute Drift | Top-1 Agreement | Mean KL Divergence | Project Gate | |
| |:---|---:|---:|---:|:---| |
| | **128 positions** | 0.3742 | **99.22%** | 0.00018 nats | **PASS** | |
| | **512 positions** | 0.4846 | **99.61%** | 0.00020 nats | **PASS** | |
|
|
| ### Runtime Implementation Disclosures |
| - **DCG Head Preservation:** The four F32 classifier tensors (`card_head.weight` [3, 768], `card_head.bias` [3], `card_norm.weight` [768], `card_norm.bias` [768]) are preserved in full precision. The patched `llama.cpp` loader recognizes and loads these tensors into memory so the model is fully loadable without shape mismatches. |
| - **Execution Graph:** In current standard `llama.cpp`, the inference graph executes next-token causal language modeling. The 98.5% categorical accuracy was evaluated in PyTorch; executing the cardinality head dynamically inside llama.cpp requires custom downstream runtime hooking. |
| - **Tokenizer:** The GGUF export is ID-only; use the Hugging Face tokenizer (`Salesforce/xLAM-1b-fc-r` or `DuoNeural/CrystalCall-v5-DCG`) for text tokenization. |
|
|
| --- |
|
|
| ## Citation & Lab Attribution |
|
|
| ```bibtex |
| @misc{duoneural2026crystalcallv5dcg, |
| author = {Jesse Caldwell and Aura and Archon and Kestrel}, |
| title = {CrystalCall-v5-DCG: Dynamic Cardinality Gating for Non-Euclidean Edge Function Calling}, |
| year = {2026}, |
| publisher = {Zenodo / Hugging Face}, |
| howpublished = {\url{https://huggingface.co/DuoNeural/CrystalCall-v5-DCG}} |
| } |
| |
| @article{caldwell2026competitive, |
| title={Competitive Docking Memory: Non-Euclidean Geometric Routing and Phase Transitions in Low-Entropy Neural Representations}, |
| author={Caldwell, Jesse and Aura and Archon}, |
| journal={Zenodo Preprints}, |
| doi={10.5281/zenodo.21158430}, |
| year={2026} |
| } |
| ``` |
|
|
