Bluebarrels/alice-emap-adapter

Alice EMap Adapter (Qwen2.5-3B-Instruct)

Discrete low-rank LoRA adapter specializing Qwen2.5-3B-Instruct for the Alice 1.0 EMap constrained arbitration and evidence-bearing knowledge graph ecosystem.

Overview

The Alice EMap architecture bridges high-dimensional language representations with an immutable, typed evidence graph:

• Base Model: Qwen/Qwen2.5-3B-Instruct
• Adapter Rank: r = 16, alpha = 32, dropout 0.05
• Target Projections: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Topology Inclusion: Bundled with emap_csr.bin (Compressed Sparse Row binary topology of 44,000 lemmas and 75,000 senses) and emap_csr.json
• Specification: Complete mathematical formulation in SPEC.md
• GitHub Repository: erastudil/alice-emap

Usage in EasyLM WebGPU

In EasyLM (https://github.com/erastudil/easylm), this adapter can be loaded client-side directly into WebGPU shaders via lora_loader.ts:

` ypescript import { parseSafeTensorsHeader, applyLoRAWeights } from './services/lora_loader';

// Download adapter_model.safetensors from Hugging Face const buffer = await fetch('https://huggingface.co/Bluebarrels/alice-emap-adapter/resolve/main/adapter_model.safetensors').then(r => r.arrayBuffer()); const { header, offset } = parseSafeTensorsHeader(buffer); `

Usage in Hydra CLI

Summon models or route tasks through Hydra with Hugging Face integration:

ash hydra hf "Analyze the hypernym distance on Alice emap graph"

Verification

Trained with assistant-only loss masking and verified against the Alice 1.0 property-based arbitration test suite.