metadata
license: mit
language:
- en
tags:
- governed-language-model
- semiconductor
- conversational
- governance
pipeline_tag: text-generation
Axiom-560M
A Governed Language Model — every output ships its own proof of governance.
Axiom-560M is a dual-mode decoder (conversational + semiconductor) trained on 56,000 governed pairs. Governance isn't a filter — it's the architecture.
Model Details
| Architecture | BLOOM-560M (decoder-only transformer) |
| Parameters | 559M |
| Training data | 56,000 governed pairs (conversational + semiconductor RTL) |
| Eval loss | 0.1635 |
| Perplexity | 1.18 overall (1.16 conversational, 1.64 semiconductor) |
| License | MIT |
Modes
Conversational — governed dialogue (perplexity 1.16)
Semiconductor — governed RTL and hardware specifications (perplexity 1.64)
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("MetaCortex-Dynamics/Axiom-560M")
tokenizer = AutoTokenizer.from_pretrained("MetaCortex-Dynamics/Axiom-560M")
input_ids = tokenizer.encode("<|conv|>What is governed generation?", return_tensors="pt")
output = model.generate(input_ids, max_new_tokens=200, temperature=0.7, do_sample=True)
print(tokenizer.decode(output[0], skip_special_tokens=True))
Governance
Every output passes through a four-phase governance pipeline:
PROPOSE → DECIDE → PROMOTE → EXECUTE
- 15 grounding operators as token vocabulary
- 7 interrogative witnesses as grammar
- Admissibility gates (G₁-G₇) with three-valued semantics
- Machine-verifiable governance trace on every output
Links
- Interactive Demo — try Axiom in your browser
- Source Code — MIT license
- Benchmark Results — 100% governance vs 0% all LLMs