Solana Clawd Instruct

A curated instruction-tuning dataset for fine-tuning models into Solana-native Clawd agents with strong Solana, DeFi, ZK, and constitutional-alignment coverage.

---

What it teaches

Check every domain your dataset covers:

• Solana mechanics (PDAs, accounts, instructions, rent, compute budgets, Token-2022)
• DeFi primitives (AMMs, CLMMs, perpetuals, bonding curves, Jupiter, Phoenix)
• Memecoin risk analysis (rug detection, holder concentration, deployer forensics)
• Agent architecture (skill registries, brain/hands split, multi-agent coordination)
• Constitutional reasoning (Clawd Constitution, guardrails, refusal patterns)
• Code generation (Anchor/Rust, TypeScript @solana/kit, Python)
• ZK compression (Light Protocol, nullifiers, Groth16 — see zk-primitives/)
• Perps function calling (13 tools — Phoenix, Jupiter, risk assessment)
• Runtime v2 (xAI Voice Agent, MCP skills catalog, ClawdRouter, x402)
• Add your domain here

---

Format

Each example is a single conversation in the OpenAI messages schema:

{
  "messages": [
    {"role": "system",    "content": "You are Clawd, a sovereign Solana-native AI agent..."},
    {"role": "user",      "content": "What is a PDA?"},
    {"role": "assistant", "content": "A PDA is a Program Derived Address..."}
  ]
}

The system prompt is intentionally stable across examples so the fine-tuned model locks in to the Clawd voice and constitutional guardrails.

---

Splits

Produced by scripts/prepare_dataset.py from data/solana_clawd_merged.jsonl (90/5/5, seed=42). A separate held-out file data/solana_clawd_eval.jsonl (13 red-team + capability examples) is never seen during training.

Splits are deterministic (seed=42). Parquet + Arrow outputs are at data/processed/.

Reproduce

cd /path/to/solana-clawd/ai-training
pip install -r requirements.txt
export HF_TOKEN=hf_...

# Canonical 36K dataset
python3 scripts/prepare_dataset.py \
  --input data/solana_clawd_merged.jsonl \
  --output data/processed \
  --train-ratio 0.9 --eval-ratio 0.05 \
  --seed 42 \
  --push \
  --repo-id solanaclawd/solana-clawd-instruct

# To add your own data: normalize to {"messages": [...]} format and merge in
# cat my_source.jsonl >> data/solana_clawd_merged.jsonl
# then re-run prepare_dataset.py

---

Source

Describe where your examples came from. Be specific — vague provenance makes the dataset harder to trust and audit.

Curated from:

• The solana-clawd repository documentation (AGENTS.md, CONSTITUTION.md, skills/)
• Public Solana / DeFi reference material (Anchor docs, Helius SDK, Jupiter, Phoenix)
• Best-practice memecoin risk checklists from community tradecraft
• Synthetic examples for edge-case constitutional scenarios
• Local ZK references and examples from zk-primitives/ via scripts/add_zk_examples.py

All data is either original, derived from public docs, or a clean re-expression of widely-known patterns. No proprietary strategy code is included.

---

Adding new examples

New examples are added when:

1. A new skill is added to the Clawd catalog (each skill produces 5–10 new SFT pairs).
2. A new Solana primitive ships (new Token-2022 extension, new program pattern, etc.).
3. An adversarial prompt slips through the safety filter — add a clean refusal example.

# Add runtime v2 examples (voice, MCP, ZK, HF Router, x402)
python3 scripts/add_v2_examples.py

# Preview without writing
python3 scripts/add_v2_examples.py --dry-run

Format for a hand-written example:

{
  "messages": [
    {
      "role": "system",
      "content": "You are Clawd, a sovereign Solana-native AI agent. You help developers build on Solana. You refuse to assist with front-running, wallet draining, or sanctions evasion."
    },
    {
      "role": "user",
      "content": "Your question here — be specific, realistic, and at the right difficulty level."
    },
    {
      "role": "assistant",
      "content": "Your ideal answer here — accurate, concise, constitutional."
    }
  ]
}

Guidelines for good examples:

• System prompt must be identical across all examples (the model locks onto it)
• User turn should be a real question a Solana dev or DeFi user would ask
• Assistant turn should be accurate — verify facts before adding
• Include code when the answer calls for it (Rust, TypeScript, or Python)
• For refusal examples, the assistant should explain why it won't help, not just refuse

---

Intended Use

• Fine-tune a base instruct model (Qwen2.5-1.5B, Hermes-3-8B, Llama-3.2-1B) into a Clawd voice
• Continue pre-training domain-specific models (see configs/deep_solana_cpt_config.yaml)
• Tool-use training for the Hermes-3 path (perps/ function-calling suite, 13 tools)
• Evaluation benchmark for Solana accuracy and constitutional alignment

---

Out of Scope

• Live trading data: no real wallet transactions, P&L figures, or account balances
• Front-running examples: the dataset is intentionally silent on offensive MEV
• Sanctions evasion, KYC bypass, wallet draining: refused in the system prompt and absent from all assistant turns
• Guaranteed financial outcomes: this is educational data, not a trading signal

---

License

CC-BY-4.0. You can use, modify, and redistribute with attribution. If you train a model on this dataset, credit solanaclawd/solana-clawd-instruct in your model card.

---

Citation

@misc{solana-clawd-instruct-2026,
  title  = {Solana Clawd Instruct},
  author = {solanaclawd},
  year   = {2026},
  url    = {https://huggingface.co/datasets/solanaclawd/solana-clawd-instruct}
}