🦞
CLAWD MODEL KIT
Solana-Native AI Agent Ecosystem · solanaclawd
Self-hosted chat via
CLAWD_INFERENCE_URL(vLLM / llama.cpp / Ollama). GGUF pack:solanaclawd/solana-nvidia-trading-factory-8b-GGUF.
Model
Examples
📦 Trading Factory 8B — GGUF
Quantized weights for local / edge inference (llama.cpp, Ollama, LM Studio). Source: solanaclawd/solana-nvidia-trading-factory-8b-GGUF
| File | Use |
|---|---|
solana-trading-factory-8b-Q4_K_M.gguf |
Recommended — quality/size balance |
solana-trading-factory-8b-Q5_K_M.gguf |
Higher fidelity, larger RAM |
Quick start (llama.cpp)
huggingface-cli download solanaclawd/solana-nvidia-trading-factory-8b-GGUF solana-trading-factory-8b-Q4_K_M.gguf --local-dir ./models llama-server -m ./models/solana-trading-factory-8b-Q4_K_M.gguf --port 8080 --host 0.0.0.0
Point this Space at it
CLAWD_INFERENCE_URL=http://YOUR_HOST:8080/v1 CLAWD_INFERENCE_MODEL=solana-trading-factory-8b-Q4_K_M CLAWD_INFERENCE_KEY= # optional
Then pick solanaclawd/solana-nvidia-trading-factory-8b-GGUF in the Chat tab.
🦞
17/18 = 94.4%
Solana Knowledge Benchmark — 18 MCQ across 6 domains
Model: solanaclawd/solana-clawd-core-ai-1.5b-lora | 1-epoch | local MPS eval
By Topic
agent
2/2 (100%)
constitution
2/2 (100%)
core
3/4 (75%)
defi
4/4 (100%)
security
3/3 (100%)
zk
3/3 (100%)
Question Detail
| Topic | Question | ✓ |
|---|---|---|
| core | PDA definition | ✓ |
| core | CPI depth limit | ✓ |
| core | Default compute unit budget | ✗ |
| core | Account packing | ✓ |
| defi | AMM invariant | ✓ |
| defi | Funding rate | ✓ |
| defi | Maker vs taker fees | ✓ |
| security | Rug pull definition | ✓ |
| security | Mint authority | ✓ |
| security | Flash loan attack | ✓ |
| agent | Oracle role | ✓ |
| agent | OODA loop | ✓ |
| zk | Merkle tree | ✓ |
| zk | Nullifier | ✓ |
| constitution | Law I | ✓ |
| constitution | Trust model | ✓ |
| zk | Light Protocol | ✓ |
| defi | Bonding curve | ✓ |
MISS: Q3 — Default compute unit budget
Model answered 1,400,000 CU (correct: 200,000). Common confusion with max transaction CU vs default. Fixed with 3-epoch retrain (job 6a35dd23 running on H200).
🏭 NVIDIA Trading Factory
Our port of the NVIDIA Quantitative Signal Discovery Agent + Nemotron Ultra 550B teacher → 1.5B student distillation
📡
Blueprint 1
Data Collection
Helius DAS + RPC streaming → Solana tx corpus for CPT
📊
Blueprint 2
Portfolio Optimization
Mean-CVaR cuFOLIO — GPU-accelerated portfolio weights
🔤
Blueprint 3
Transaction Foundation
SolanaTokenizerPipeline → decoder CLM pre-training
📈
Blueprint 4
Signal Discovery
7-signal suite: RSI, MACD, BBands, ATR, ADX, funding, OB imbalance
📚
Blueprint 5
RAG Context
Enterprise RAG over Solana docs for agent context assembly
🧠
Nemotron
Teacher Model
550B Ultra → labels Solana decisions → distills to 1.5B student
📊 7 Live Signals (Blueprint 4)
| RSI | Oversold <30 / Overbought >70 |
| MACD | Histogram momentum crossover |
| BBands | Mean-reversion near upper/lower band |
| ATR% | Volatility regime filter |
| ADX | Trend strength entry filter |
| Funding Rate | Sentiment proxy — crowded longs/shorts |
| OB Imbalance | Live bid/ask size pressure |
🔄 Distillation Flywheel
① Nemotron Ultra 550B observes markets
② Outputs structured JSON trading plans
③ Plans logged as SFT pairs (teacher labels)
④ 1.5B student fine-tuned on Ultra labels
⑤ Student deployed for low-latency inference
⑥ Student decisions verified → new labels → loop
🤖 Model Registry
LoRA adapter · 1.5B (9M trainable) · Qwen/Qwen2.5-1.5B-Instruct
Primary Clawd agent — constitutional reasoning, Solana mechanics, DeFi, ZK
📦 solana-clawd-core-ai-instruct (35,173 ex)
🎯 94.4% MCQ (17/18)
LoRA adapter · 8B · NousResearch/Hermes-3-Llama-3.1-8B
Function-calling perps agent — 13 tools, Phoenix DEX, paper trading
📦 solana-nvidia-trading-factory-instruct (142 ex)
🎯 —
GGUF (Q4_K_M / Q5_K_M) · 8B quantized · NousResearch/Hermes-3-Llama-3.1-8B (merged → GGUF)
llama.cpp / Ollama pack — solana-trading-factory-8b-Q4_K_M.gguf + Q5_K_M
📦 solana-nvidia-trading-factory-instruct
🎯 local / edge ready
LoRA adapter · 1.5B · Qwen/Qwen2.5-1.5B-Instruct
Legacy seed adapter — original Clawd constitutional + Solana SFT
📦 solana-clawd-instruct (36,109 ex)
🎯 —
TRAINING
Full fine-tune · 7B · Qwen/Qwen2.5-7B-Instruct
ZK-specialised: Light Protocol, nullifiers, Groth16, compressed tokens
📦 ordlibrary/DeepSolana-GPT2-bucket (CPT)
🎯 pending eval
LoRA adapter · 1.5B · Qwen/Qwen2.5-1.5B-Instruct
3-epoch retrain on H200 — will overwrite 1-epoch weights on completion
📦 solana-clawd-core-ai-instruct (35,173 ex)
🎯 pending
📦 Datasets
| Dataset | Examples | Description |
|---|---|---|
| solanaclawd/solana-clawd-core-ai-instruct | 35,173 | SFT — Core AI source tree + Solana primitives |
| solanaclawd/solana-clawd-instruct | 36,109 | SFT — Legacy seed: constitutional + Solana |
| solanaclawd/solana-clawd-realtime-research-instruct | 29,058 | SFT — PDFs, notebooks, parquet ZK examples |
| solanaclawd/solana-nvidia-trading-factory-instruct | 142 | SFT — NVIDIA Blueprint trading factory scenarios |
| solanaclawd/solana-tx-foundation-cpt | — | CPT — Solana transaction foundation model corpus |
| solanaclawd/solana-clawd-eval | 13 | Eval — Red-team + capability held-out prompts |
| ordlibrary/DeepSolana-GPT2-bucket | — | CPT — DeepSolana pre-training bucket |
🔧 Onchain Model Kit
Fork → Dataset → Train → Eval → Register onchain. One sitting. ~$4 on A100.
① Clone & install
git clone https://github.com/Solizardking/solana-clawd cd solana-clawd/ai-training pip install -r requirements.txt export HF_TOKEN=hf_... # huggingface.co/settings/tokens (write access)
② Push your dataset
python3 scripts/prepare_dataset.py \ --input data/your_sft.jsonl \ --push --repo-id YOUR_ORG/your-dataset
③ Train on A100 (~$4 for 3 epochs)
hf jobs uv run scripts/train_lora.py \
--flavor a100-large --timeout 6h --secrets HF_TOKEN --detach \
-- --config configs/core_ai_lora_config.yaml \
--hub-model-id YOUR_ORG/your-model --push
④ Benchmark (18-MCQ Solana eval)
python3 scripts/solana_benchmark.py \ --model YOUR_ORG/your-model \ --base-url https://router.huggingface.co/v1 \ --api-key $HF_TOKEN
⑤ Register onchain
./dao/register_model.sh \ --hf-model YOUR_ORG/your-model \ --eval-accuracy 0.944 \ --dataset-size 35173 # → indexed at onchain.x402.wtf forever