easylm-gemma-4-e2b-it

EasyLM Fine-Tuned LoRA Adapter for google/gemma-4-E2B-it

Hey, here are the base models we used, and here is what we trained them on:

• Base Model: google/gemma-4-E2B-it
• EasyLM Target Tier: 2GB Ultralight / Mobile
• Runtime Footprint: ~1.2 GB weights | 4,096 tokens context window
• Training Dataset: Bluebarrels/easylm-training-corpus

What We Trained This Model On

This adapter was trained on the official EasyLM sovereign instruction-tuning corpus:

1. 32 Sovereign Academic Stacks: Mathematics, physics, chemistry, biology, computing, software engineering, law, philosophy, history, and economics structured in Progen topic:comment dialect.
2. EasyLM Core Features: AtMem (local zero-vector IndexedDB atomic memory), ZCABS (Zero-Correlation Anti-Bullshit canary nonce system), and Anti-Loop Protection.
3. Hands Tool Execution: Structured tool-calling for live web search, Python calculation, bash terminal commands, and Alice Cognitive Mind epistemic queries.
4. Hardware Adaptation: Client-side context budgeting, mobile memory fences (<1GB iOS budget), and Kid Safe boundary guardrails.

Hyperparameters

• Method: Low-Rank Adaptation (LoRA)
• Rank ($r$): 16
• Alpha ($lpha$): 32
• Dropout: 0.05
• Target Projection Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Precision: bfloat16 mixed precision
• Training Framework: TRL SFTTrainer on Modal serverless GPU infrastructure

Client-Side WebGPU Usage

import { CreateMLCEngine } from '@mlc-ai/web-llm';

const engine = await CreateMLCEngine('google/gemma-4-E2B-it', {
  appConfig: {
    model_list: [
      {
        model: 'https://huggingface.co/google/gemma-4-E2B-it',
        model_id: 'easylm-gemma-4-e2b-it',
        model_lib: 'https://raw.githubusercontent.com/mlc-ai/binary-mlc-llm-libs/main/web-llm-models/v0_2_84/base/Qwen2-7B-Instruct-q4f16_1_cs1k-webgpu.wasm',
        vram_required_MB: 2600
      }
    ]
  }
});