Model card & methodology

Model

• Name: sovereign-v2:v6-sft
• Base: Qwen3-14B
• Size: 14B parameters
• Quantization: Q6_K GGUF for deployment (12 GB)
• Served via: Ollama on V100 + RTX 2070

Training

• Method: Supervised Fine-Tuning (SFT) with curriculum weighting. DPO tried and rolled back (regressed).
• Dataset: 2,500 records mixed across trading categories — quant, DeFi, MEV, Solana, L2s, tokenomics, security.
• Framework: Unsloth (LoRA r=32, lr=1e-5, 2 epochs, curriculum-weighted weak categories)
• Compute: V100 16GB for ~6 hours. Merge + GGUF export separate step.

Benchmark

• 45 questions, 8 categories, A–F grading.
• Graded independently by two reviewers plus rubric matcher; conflicts resolved by a third reviewer.
• Current score: 87% A-rate, 98% A+B.
• Methodology + dataset: /benchmarks

Known weaknesses

• Smart-money / wallet clustering analysis (graded C) — pre-train data underrepresented here. Working on it.
• Regime detection (graded B) — lacks explicit time-series intuition without extra context.
• Zero knowledge of events after training cutoff (Apr 2026).
• Cannot execute live on-chain queries — requires you to provide current context in the prompt.

Privacy

API requests are logged with user ID, prompt hash, and latency metrics. Prompts themselves are not stored long-term; they pass through Ollama memory only during inference. Full privacy details at /legal.

Versioning

• Every model version gets a full public benchmark before promotion to :latest.
• Old versions remain accessible via explicit tag (e.g. sovereign-v2:v5).
• Change log is published with each promotion.

Get involved

The eval set is open-source. PRs adding new questions (with rubrics) are welcome. Community channels.