solanaclawd/README

Solana Clawd

The Hugging Face home for the Solana Clawd model stack: public-safe datasets, LoRA adapters, evaluation artifacts, and CAAP/1.0 registry metadata for Solana-native AI agents.

GitHub: Solizardking/solana-clawd
Onchain registry: onchain.x402.wtf
Registry JSON: /.well-known/clawd-registry.json
Model kit: ai-training/model-kit

Solana AI Model Kit

The kit is a one-shot path for building, publishing, training, registering, and serving Solana AI models.

# Safe default: clone/update the repo, audit local release state, print next steps.
curl -fsSL https://raw.githubusercontent.com/Solizardking/solana-clawd/main/ai-training/scripts/solana_ai_model_kit.sh | bash

# From a checkout:
npm run model-kit
npm run model-kit:register
npm run model-kit:train

Live CAAP/1.0 registry POST:

bash ai-training/scripts/solana_ai_model_kit.sh \
  --local \
  --live-register \
  --hf-model YOUR_ORG/your-model \
  --endpoint https://your-router.example/v1 \
  --eval-accuracy 0.60 \
  --dataset-size 35173

Current Artifacts

Datasets

Repo	Examples	What is inside
solanaclawd/solana-clawd-core-ai-instruct	35,173	Public-safe blend of core-ai, Helius/Clawd runtime files, knowledge JSONL, and cleaned SFT examples
solanaclawd/solana-clawd-realtime-research-instruct	29,058	PDFs, notebooks, parquet Solana QA, ZK skill context, and realtime document ingestion outputs
solanaclawd/solana-clawd-nvidia-trading-factory-instruct	142	NVIDIA trading-factory plans, Solana spot/perps scenarios, cuFOLIO/cuOpt handoffs, Phoenix/Vulcan paper strategies, Rise read plans, and risk refusals
solanaclawd/solana-clawd-eval	13	Held-out capability, calibration, and red-team prompts

Models

Repo	Status	Base
solanaclawd/solana-clawd-core-ai-1.5b-lora	Recovery job ordlibrary/6a35a6833093dba73ce2a86b is running on a100-large; first HF job trained then failed during Hub push	Qwen/Qwen2.5-1.5B-Instruct
solanaclawd/solana-nvidia-trading-factory-8b-lora	Completed HF job ordlibrary/6a35a2ce953ed90bfb945009; train loss 1.1692, eval loss 0.8064, eval token accuracy 0.8547	NousResearch/Hermes-3-Llama-3.1-8B
solanaclawd/solana-clawd-1.5b	Merged-model target	Qwen2.5 1.5B + LoRA
solanaclawd/solana-clawd-7b-lora	Optional larger target	Qwen2.5 7B

Training Status

• Active Core AI retry: ordlibrary/6a35a6833093dba73ce2a86b
• Core recovery evidence: loaded solanaclawd/solana-clawd-core-ai-instruct, tokenized all 31,655 train rows, entered training, and reached at least step 221/3957 with mean token accuracy around 0.79.
• Superseded failed trading job: ordlibrary/6a359f0e953ed90bfb944faf
• Failure mode: the HF job tried to load /data/nvidia_trading_factory_processed instead of the published Hub dataset.
• Fix: scripts/train_lora.py now falls back to dataset_repo when the configured local path is absent, and prepare_dataset.py normalizes metadata across train/eval/test splits for Hub pushes.
• Superseded failed replacement: ordlibrary/6a35a02d953ed90bfb944fe3
• Second fix: Hermes exposes tokenizer.chat_template as a dict and TRL expected a string when assistant-only loss was enabled. The trainer now normalizes dict templates and disables assistant-only loss when generation markers are unavailable.
• Successful retry: ordlibrary/6a35a2ce953ed90bfb945009
• Final evidence: active retry loaded the published Hub dataset, tokenized train/eval splits, built SFTTrainer, completed 48/48 steps, pushed adapter files, and verified adapter_config.json plus adapter_model.safetensors on Hub.
• Final metrics: train loss 1.1692, eval loss 0.8064, eval mean token accuracy 0.8547.
• W&B: disabled unless WANDB_API_KEY is present in the launching environment.

Onchain Registry

The registry API is served by OnChain-AI and indexed at onchain.x402.wtf.

curl -sS https://onchain.x402.wtf/.well-known/clawd-registry.json | python3 -m json.tool
curl -sS "https://onchain.x402.wtf/api/models?hf_id=solanaclawd/solana-clawd-core-ai-1.5b-lora" | python3 -m json.tool

Local sidecar:

export ONCHAIN_AI_ROOT=/Users/8bit/Downloads/OnChain-Ai-main

cd "$ONCHAIN_AI_ROOT/backend"
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install -r requirements.txt
PORT=5001 python3 main.py

cd "$ONCHAIN_AI_ROOT/frontend"
npm install
VITE_API_BASE_URL=http://localhost:5001 npm run dev

Safety

• No private keys, API tokens, OAuth client secrets, Google ADC JSON, W&B keys, or HF tokens belong in datasets, cards, commits, manifests, or Hub uploads.
• Trading-factory data defaults to paper mode.
• Live execution belongs outside model training data and requires explicit operator approval, wallet isolation, and pre-trade risk checks.
• The model is the planning layer; key-bearing execution clients are separate trust domains.

Links

• Solana Clawd GitHub
• AI training README
• Model kit guide
• Dataset card
• Model card