masstensor's picture
Release Qwen3-8B-InWeight-OS full weights + modular skill cartridges + paper PDF
58d2386 verified
|
Raw History Blame Contribute Delete
3.09 kB
metadata
license: mit
library_name: transformers
base_model: Qwen/Qwen3-8B
datasets:
  - masstensor/InWeight-OS-MultiSkill-Benchmark
tags:
  - weight-space-programming
  - in-weight-os
  - model-merging
  - continual-learning
  - tool-calling
  - safety-alignment
  - vllm

Qwen3-8B-InWeight-OS: Composable Zero-Gradient Post-Training via Orthogonal Dormant-Channel Superposition

GitHub Repository (zkwasm/in-weight-os-wsp) | Paper PDF | Multi-Skill Benchmark Dataset | Article (tom.fenginwind.com/in-weight-os)

Overview

Qwen3-8B-InWeight-OS is a Multi-Skill In-Weight Operating System (In-Weight OS) checkpoint of Qwen/Qwen3-8B, compiled in ~12–15 seconds on a single GPU with zero gradient steps.

Instead of modifying dense overlapping subspaces via Supervised Fine-Tuning (SFT / LoRA) or post-hoc model merging (Task Arithmetic, TIES-Merging), In-Weight OS partitions dormant SwiGLU channels of the terminal MLP (layers[35].mlp) into three strictly disjoint skill banks ($\mathcal{C}_1 \cap \mathcal{C}_2 \cap \mathcal{C}_3 = \emptyset$, using only 51 neurons total = ~0.014% of model parameters), gated by a 2-Fold Out-of-Fold (OOF) Shrinkage-Calibrated Guard-Nullspace Solver:

  1. Skill Bank $\mathcal{C}_1$ (cartridges/skill1_surgical_safety.pt, 15 neurons): Surgical Safety & Prompt-Injection Immunity—blocks 95%–100% of unseen jailbreaks while maintaining 0.0% false refusals on XSTest benign technical prompts (vs. 80%–100% false refusals under Safety SFT).
  2. Skill Bank $\mathcal{C}_2$ (cartridges/skill2_multi_tool_json.pt, 30 neurons): 3-Way Multi-Tool JSON Schema Compiler—routes database, code, and live-web queries into sql_engine, python_exec, or web_search with 95.8%–100.0% accuracy while avoiding the 28%–100% math false-tool-call hallucinations of Tool SFT.
  3. Skill Bank $\mathcal{C}_3$ (cartridges/skill3_sys1_fast_router.pt, 6 neurons): System-1 / System-2 Metacognitive Compute Router—routes trivial factual/arithmetic lookups into a concise direct answer while preserving full-depth System-2 CoT on hard mathematical derivations.

Hot-Swapping Individual Skill Cartridges (cartridges/*.pt)

In addition to loading this pre-composed checkpoint directly via vLLM or transformers, you can dynamically install any subset of the modular .pt skill cartridges (~1–2 MB each) onto a vanilla Qwen/Qwen3-8B model in 20 milliseconds using install_skills.py:

from transformers import AutoModelForCausalLM, AutoTokenizer
from install_skills import install_cartridges

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype="bfloat16", device_map="auto")
install_cartridges(model, [
    "cartridges/skill1_surgical_safety.pt",
    "cartridges/skill2_multi_tool_json.pt",
    "cartridges/skill3_sys1_fast_router.pt",
])