What We Do

Most post-training in the open-weight space uses naive abliteration — subtracting a refusal direction from model weights. This works until it doesn't: it destroys capability-entangled representations, causing 8+ point drops on TruthfulQA and high KL divergence from the base model.

We do things differently.

SABER (Spectral Analysis-Based Entanglement Resolution) uses Canonical Correlation Analysis to identify which latent directions encode refusal vs. which encode useful capabilities, then surgically edits only the refusal-correlated subspace. The result: models that refuse harmful requests while retaining reasoning, coding, and factual accuracy.

NSC-ACE (Neural Steering Committee for Agentic Co-Evolution) is our GRPO-based RL method that trains steering vectors alongside policy optimization, producing models with better instruction-following and latent controllability.

Model Lines

NSC-ACE-SABER Pipeline

Our flagship post-training pipeline. Full SFT → NSC-ACE (GRPO with latent steering) → SABER → GGUF quantization.

→ View Collection

Ornstein Series

Post-trained Ornstein models: base, Hermes-tuned, and SABER-processed. Available in GGUF (llama.cpp/ollama) and MLX (Apple Silicon).

→ View Collection

BOREAL

From-scratch pretraining project. DeltaNet hybrid architecture with DeepSeek-V4 routing and Temporal Shift Tokens. Currently in early stages — 250M proof-of-concept, with 2B and 10B-MoE planned.

→ View Collection

BusyBeaver

Tool-policy research models for agentic AI safety. Studying how small models learn to follow tool-use policies and refuse unsafe tool calls.

→ View Collection

Methods

Philosophy

• Open artifacts — full weights, configs, and training code. Inspect everything.
• Local-first — every release includes GGUF and/or MLX quantizations for local inference.
• Capability-preserving — we measure what we break. If SABER drops a benchmark, we iterate.
• No naive abliteration — direction subtraction is a blunt instrument. We use spectral methods.

Contact

Open a discussion on any model repo, or reach the founder at DJLougen.