meshscale-worker-template / MODELS_TO_PLUG.md
tostido's picture
Build MeshScale CPU worker template
96ef23c verified
|
Raw
History Blame Contribute Delete
6.77 kB

MODELS TO PLUG - ARC-AGI 48-SLOT ARCHITECTURE

Ready-to-use HuggingFace model IDs for relay system


TIER 1: PERCEPTION BANK (8 slots)

Slot 0: SigLIP (Patch-level vision)

hub search google/siglip-base-patch16-224
1

Model: google/siglip-base-patch16-224
VRAM: ~0.8 GB

Slot 1: CLIP (Global vision)

hub search openai/clip-vit-large-patch14
1

Model: openai/clip-vit-large-patch14
VRAM: ~0.8 GB

Slot 2: DINOv2 (Self-supervised)

hub search facebook/dinov2-base
2

Model: facebook/dinov2-base
VRAM: ~0.6 GB

Slot 3: SAM Encoder (Segmentation)

hub search facebook/sam-vit-base
1

Model: facebook/sam-vit-base
VRAM: ~1.2 GB

Slot 4: DETR (Object detection)

hub search facebook/detr-resnet-50
1

Model: facebook/detr-resnet-50
VRAM: ~0.1 GB

Slots 5-7: Custom models (TBD)

  • Edge detector (custom CNN)
  • Color quantizer (custom)
  • Grid parser (custom transformer)

TIER 1 TOTAL: ~3.7 GB


TIER 2: PRIMITIVE SPECIALISTS (16 slots)

Strategy: Fine-tune small models (DistilBERT, Small ViT) for each primitive

Base models to fine-tune:

hub search distilbert-base-uncased
hub search google/vit-base-patch16-224
hub search microsoft/resnet-50

Per specialist: ~0.3B params, ~0.6 GB VRAM
TIER 2 TOTAL: ~9.6 GB


TIER 3: PROGRAM SYNTHESIS (8 slots)

Slot 24: DeepSeek Coder 33B

hub search deepseek-ai/deepseek-coder-33b-base
44

Model: deepseek-ai/deepseek-coder-33b-base
VRAM: 66 GB (FP16) / 16.5 GB (INT4)

Slot 25: CodeLlama 34B

hub search codellama/CodeLlama-34b-hf
1

Model: codellama/CodeLlama-34b-hf
VRAM: 68 GB (FP16) / 17 GB (INT4)

Slot 26: StarCoder2 15B

hub search bigcode/starcoder2-15b
1

Model: bigcode/starcoder2-15b
VRAM: 30 GB (FP16) / 7.5 GB (INT4)

Slot 27: Qwen2-Coder 32B

hub search Qwen/Qwen2.5-Coder-32B-Instruct
1

Model: Qwen/Qwen2.5-Coder-32B-Instruct
VRAM: 64 GB (FP16) / 16 GB (INT4)

Slot 28: DSL Generator (Fine-tuned CodeLlama 7B)

hub search codellama/CodeLlama-7b-hf
1

Model: codellama/CodeLlama-7b-hf (to be fine-tuned)
VRAM: 14 GB (FP16) / 3.5 GB (INT4)

Slot 29: Tree Builder (Custom)

TBD - Custom transformer for AST generation

Slot 30: Genetic Programmer

hub search mistralai/Mistral-7B-v0.1
1

Model: mistralai/Mistral-7B-v0.1
VRAM: 14 GB (FP16) / 3.5 GB (INT4)

Slot 31: Analogy Mapper

hub search BAAI/bge-large-en-v1.5
1

Model: BAAI/bge-large-en-v1.5
VRAM: 2 GB

TIER 3 TOTAL: 264 GB (FP16) / 66 GB (INT4)


TIER 4: REASONING & INDUCTION (6 slots)

Slot 32: Llama 3.1 70B

hub search meta-llama/Llama-3.1-70B-Instruct
1

Model: meta-llama/Llama-3.1-70B-Instruct
VRAM: 140 GB (FP16) / 35 GB (INT4)

Slot 33: Claude-distill (Proxy with smaller model)

hub search mistralai/Mistral-7B-Instruct-v0.3
1

Model: mistralai/Mistral-7B-Instruct-v0.3
VRAM: 14 GB (FP16) / 3.5 GB (INT4)

Slot 34: Qwen2-Math 72B

hub search Qwen/Qwen2.5-Math-72B-Instruct
1

Model: Qwen/Qwen2.5-Math-72B-Instruct
VRAM: 144 GB (FP16) / 36 GB (INT4)

Slot 35: Gemma 2 27B

hub search google/gemma-2-27b-it
1

Model: google/gemma-2-27b-it
VRAM: 54 GB (FP16) / 13.5 GB (INT4)

Slot 36: Rule Inducer (Fine-tuned Llama 13B)

hub search meta-llama/Llama-2-13b-hf
1

Model: meta-llama/Llama-2-13b-hf (to be fine-tuned)
VRAM: 26 GB (FP16) / 6.5 GB (INT4)

Slot 37: Hypothesis Generator

hub search mistralai/Mistral-7B-Instruct-v0.3
1

Model: mistralai/Mistral-7B-Instruct-v0.3
VRAM: 14 GB (FP16) / 3.5 GB (INT4)

TIER 4 TOTAL: 392 GB (FP16) / 98 GB (INT4)


TIER 5: VERIFICATION (4 slots)

Slots 38-41: Lightweight verification tools

  • Python executor (built-in)
  • Similarity scorer (numpy/sklearn)
  • Z3 solver (symbolic verification)
  • Confidence scorer (small classifier)

TIER 5 TOTAL: ~1.8 GB


TIER 6: SEARCH & META (4 slots)

Slot 44: Task Classifier

hub search distilbert-base-uncased
1

Model: distilbert-base-uncased (to be fine-tuned)
VRAM: ~0.5 GB

Slots 42-43, 45: Custom algorithms

  • MCTS controller (Python)
  • Beam ranker (small scorer)
  • Difficulty estimator (regressor)

TIER 6 TOTAL: ~2.3 GB


TIER 7: MEMORY & RETRIEVAL (2 slots)

Slot 46: Task Embedder

hub search BAAI/bge-large-en-v1.5
1

Model: BAAI/bge-large-en-v1.5
VRAM: ~0.6 GB

Slot 47: Solution Bank

  • FAISS vector database
  • Populated with 400 ARC training tasks

TIER 7 TOTAL: ~2.6 GB


QUICK START SEQUENCE (After GPU rental)

1. Load Core Infrastructure (< 10 GB)

# Perception
hub search google/siglip-base-patch16-224
1
hub search openai/clip-vit-large-patch14
1
hub search facebook/dinov2-base
2

# Memory
hub search BAAI/bge-large-en-v1.5
1

# Task classifier
hub search distilbert-base-uncased
1

2. Load Reasoning Model (+ 35 GB)

# Llama 3.1 70B (INT4 quantized)
hub search meta-llama/Llama-3.1-70B-Instruct
# Look for INT4/GPTQ version

3. Load Synthesis Model (swap with reasoning)

# DeepSeek Coder 33B (INT4 quantized)
hub search deepseek-ai/deepseek-coder-33b-base
44

4. Test on ARC Task

# Load ARC task
infer "Analyze this grid pattern..."

# Run through pipeline
deliberate "What transformation is needed?"

QUANTIZED MODEL SEARCH TIPS

For INT4/GPTQ versions (4× smaller VRAM):

hub search meta-llama/Llama-3.1-70B GPTQ
hub search meta-llama/Llama-3.1-70B INT4
hub search deepseek-ai/deepseek-coder-33b GPTQ

Look for model IDs containing:

  • GPTQ
  • INT4
  • 4bit
  • bnb-4bit (bitsandbytes)
  • GGUF (for llama.cpp compatibility)

RELAY COMMANDS CHEAT SHEET

# Search for model
hub search <query>

# Plug model by number from search results
<number>

# Plug model directly
plug <model_id>

# Plug to specific slot
plug slot N <model_id>

# Check what's loaded
who

# Check slots
slots

# Test model
embed "test text"

# Run inference
infer "input text"

# Council deliberation
deliberate "question"

# Dreamer imagination
imagine 10

All models verified on HuggingFace Hub as of 2026-01-28
Ready for ARC-AGI Prize competition