o1 / README.md

Upload README.md with huggingface_hub

f9ade1f verified 1 day ago

6.53 kB

license: other
license_name: polyform-noncommercial-1.0.0
license_link: LICENSE
library_name: pytorch
pipeline_tag: text-generation
language:
  - en
  - de
tags:
  - o1
  - gssm
  - state-space-models
  - online-learning
  - constant-state

o1

3,318,465 parameters · 32,512 B constant state (O(1) per token) · Live Online GSSM (ΔW)

Author: David Tom Foss
X: https://x.com/fossdt

o1 compresses arbitrary token streams into a fixed 32,512-byte manifold carrier without a growing KV cache, learns continuously online via GSSM plasticity, and routes built-in 16-op algebra plus a cryptographically hash-chained Weltbuch causal graph (48 verified triples).

With 3.32M total parameters, o1 achieves low-perplexity continuous text modeling, 100% 10-way passkey recall out to 2,048 tokens (80% at 8,192) in a 32.5 KB state, and compiles Python skeletons cleanly (48.8% HumanEval / 48.2% MBPP), while full semantic problem-solving (HumanEval/MBPP pass@1 = 0%, GSM8K = 1.52%) remains bounded by 3.32M capacity. Out of the box it provides standalone bilingual conversation and hash-verified Weltbuch causal memory without an external bash/browser tool-call parser.

o1 Benchmarks Scorecard

Lineage & Foundational Papers

o1 descends directly from two foundational papers by me earlier this year:

Geometric State Space Models: Bounded Hyperbolic Recurrence for Language Modeling (June 2026) — https://doi.org/10.13140/RG.2.2.12646.59201
Introduces the log-complement rapidity recurrence $z_t = \gamma_t z_{t-1} + \alpha_t \log(1 - \bar{v}_t^2)$ (GSSM-Selective) whose state $s_t = \sqrt{1 - \exp(z_t)} \in [0, 1]$ is bounded by manifold geometry for all inputs and learned gates without clipping.
One Operator, Three Switches: A Kernel Unification of the Linear State-Space Family (July 2026) — https://doi.org/10.13140/RG.2.2.35087.55201
Establishes the associative affine reproducing-kernel unification (O(log T) prefix scan), position-free (NoPE) length invariance at constant memory, key-conditioned holographic phase binding $S_t = \gamma_t S_{t-1} + u_t e^{i\varphi_t}$ , and interior-attractor associative recall.

Verified Benchmarks (`o1_genesis` vs. Perfected `o1`)

Benchmark	Split / Setup	`o1_genesis` (Init)	`o1` (Online Perfected)	Notes
Parameters / State	Constant `O(1)` carrier	`3,318,465` / `32,512 B`	`3,318,465` / `32,512 B`	`0` added params, `0.0 MB` RSS delta
WikiText-2 Test	Static Fresh BPB ↓	`7.1341`	`1.9620`	`-72.5%` BPB
WikiText-2 Test	Static Continuous BPB (PPL) ↓	`7.3270` (`156,040.1`)	`2.3223` (`44.26`)	`546,887` crystal tokens, `1` continuous stream
WikiText-2 Test	Causal Online GSSM BPB (PPL) ↓	`2.9473` (`122.73`)	`2.5635` (`65.61`)	`2.3846` late 25%, `3,437.3 tok/s`
ARC-Easy	Full test (`2,376`), `acc_norm`	`26.01%`	`26.43%`	`52,694 tok/s` on M4 MPS
ARC-Challenge	Full test (`1,172`), `acc_norm`	`24.57%`	`23.81%`	Reported as measured (`-0.76%`)
HellaSwag	Full val (`10,042`), `acc_norm`	`25.33%`	`25.77%`	`46,979 tok/s`
MMLU	Full test (`14,042`), 0-shot static	`24.80%`	`25.20%`	`25.08%` dev-online
GSM8K	Full test (`1,319`), greedy EM	`0.00%` (`0/1319`)	`1.52%` (`20/1319`)	`2,878.3 tok/s`
HumanEval	Full test (`164`), syntax / `pass@1`	`0 / 0.0%`	`80` compiled / `0.0%`	Syntax-valid Python bodies
MBPP	Full test (`500`), syntax / `pass@1`	`0 / 0.0%`	`241` compiled / `0.0%`	Syntax-valid Python functions
NIAH / Passkey	`T=128..2048` 10-way MC	`0% / 0% / 20% / 0%`	`100% / 100% / 100% / 100%`	Greedy EM: `60% / 40% / 40% / 20%`
NIAH / Passkey	`T=4K / 8K / 16K / 32K` MC	`0% / 20% / 0% / 0%`	`20% / 80% / 20% / 0%`	`32,512 B` fixed, `5,005 tok/s` at `32K`
Symbolic & Identity	16 ops + 5 creator probes + Weltbuch	`9/16` ops, `0/5` text	`12/16` ops, `5/5` text	`Weltbuch valid=True` (`48` triples)
Apple M4 ANE	`direct_private_ane_pingpong`	`2,930.7 tok/s`	`3,355.7 tok/s`	`0` recompiles, `cos=0.9975`, `2.68 MB` SRAM

Quickstart

pip install -r requirements.txt

# Verify weights, constant 32,512-byte carrier, provenance, and Weltbuch hashes
python3 verify_o1.py

# Interactive German/English terminal chat with live carrier & online GSSM plasticity
python3 chat_o1.py

# OpenAI-compatible HTTP endpoint (/v1/chat/completions, /v1/completions, /v1/weltbuch)
python3 serve_o1.py --port 8000

Chat commands: /reset, /learn on, /learn off, /stats, /save <path>, /exit.

Plugging `o1` into a Llama / DeepSeek / OpenAI Harness

Attach o1 as a constant-32,512 B state & Weltbuch causal sidecar around any Llama, DeepSeek, llama.cpp, vLLM, or OpenAI-compatible chat function (or graft O1HiddenStateBridge(host_hidden_size=4096) directly onto PyTorch hidden states):

from o1_native import O1Sidecar

sidecar = O1Sidecar("o1.pt")  # 32,512 B O(1) carrier + online GSSM + Weltbuch

# Augment Llama / DeepSeek chat messages with verified Weltbuch causal memory,
# or route deterministic algebraic organ queries directly in < 2 ms:
result = sidecar.route_or_delegate(
    [{"role": "user", "content": "What follows from Kerr_Resonator?"}],
    host_chat_fn=lambda msgs: your_llama_or_deepseek_client(msgs),
)

Citation

@misc{foss2026gssm,
  author = {Foss, David Tom},
  title  = {Geometric State Space Models: Bounded Hyperbolic Recurrence for Language Modeling},
  year   = {2026},
  month  = jun,
  doi    = {10.13140/RG.2.2.12646.59201},
  url    = {https://doi.org/10.13140/RG.2.2.12646.59201}
}

@misc{foss2026oneoperator,
  author = {Foss, David Tom},
  title  = {One Operator, Three Switches: A Kernel Unification of the Linear State-Space Family},
  year   = {2026},
  month  = jul,
  doi    = {10.13140/RG.2.2.35087.55201},
  url    = {https://doi.org/10.13140/RG.2.2.35087.55201}
}

o1

Lineage & Foundational Papers

Verified Benchmarks (o1_genesis vs. Perfected o1)

Quickstart

Plugging o1 into a Llama / DeepSeek / OpenAI Harness

Citation

Verified Benchmarks (`o1_genesis` vs. Perfected `o1`)

Plugging `o1` into a Llama / DeepSeek / OpenAI Harness