A frozen foundation,
reproducible numbers.
A five-million-parameter semantic foundation, trained and served on consumer CPUs. Weights, inference code and benchmark scripts are all in the repository.
Parameters
5,059,040
Weight size
~20 MB
End-to-end latency
100 – 200 ms
Training hardware
Consumer CPU
Latency measured on an i7-12700H with PyTorch 2.0 and seq_len 256
What it is
A frozen semantic foundation: Chinese text in, structured semantic representation out.A frozen semantic foundation: Chinese text in, structured semantic representation out.
- Input
- Chinese natural-language text, tokenised by the bundled tokeniser, sequence length 256.
- Output
- An encoding-space semantic vector, a Luoshu state vector, and a 64-dimensional hexagram prototype vector.
- Architecture
- A bilateral ladder network, pretrained on general corpora and classical Yijing texts.
- Paradigm
- Frozen foundation plus lightweight adaptation. Core weights stop changing once converged; new knowledge only enters adapters.
In the repository
- Model weights and SHA256 checksums (weights/)
- Inference interface (daoti/inference.py)
- Benchmark script (scripts/eval_benchmark.py)
- Whitepaper and 6 working papers (docs/)
- Colab demo notebook (notebooks/)
Not in the repository
- Full training code
- Data recipes
- Architecture design details and intent
Those three require separate authorisation by email. Weights are governed by the DaoTi Research License v1.0; code is Apache 2.0.
An observed phenomenon
The core projection layer converges to an almost flat solution space. We call it the degenerate ground state.The core projection layer converges to an almost flat solution space. We call it the degenerate ground state.
F(d) ∝ d⁰
Retrieval loss along text_proj is independent of the distance d in that direction — multiple equivalent solutions coexist, with different parameters but identical behaviour. That is exactly why a frozen foundation still works.
- V91
- Five paths, one destination. Five random initialisations of text_proj all reach top-1 100%, yet the mean CKA between the two representation sets is only 0.267.
- V90
- Pushed away, never returns. The regression rate is negative. The loss landscape is almost completely flat along text_proj — changing it or not lands in the same basin.
The boundary of this claim
It is confirmed only at the few-million-parameter scale. Whether it holds for larger models is unverified — that is the core risk to whether the paradigm generalises, and section 7.4 of the whitepaper says the same.
What it can do
Six capabilities, each backed by an interface you can run from the repository.Six capabilities, each backed by an interface you can run from the repository.
- Eight structured outputs
- Eight palaces, six relations, six spirits, heavenly stems, earthly branches, vigour, moving lines and five elements — all from a single forward pass.
- Three divination methods
- traditional (Zhouyi), meihua (Meihua Yishu) and liuyao (Liuyao) switch method heads on the same weights.
- RAG self-calibration
- Uses the structured result as a retrieval key; declares uncertainty when coherence drops below 0.3.
- Domain adapters
- Freeze the core, train LoRA only. Your own JSONL data extends it to a new domain without touching DaoTi.
- Physical parameter adaptation
- An external input encoder and regression head map physical parameters to spectra while the core stays frozen.
- Built-in architectural constraints
- A domain classifier up front, hexagram entropy for ambiguity, coherence gating for stray signals — no external rules.
Eight symbolic reasoning tasks
99.96%
Mean accuracy · whitepaper 5.2
Domain classification
98.3%
8 domains overall · whitepaper 5.3
Frozen adaptation V57
98.0%
Converged in 8 epochs · whitepaper 5.4
These are the whitepaper's self-reported numbers. Running python scripts/eval_benchmark.py writes benchmarks/benchmark_results.json, which contains eight-palace classification, 64-hexagram retrieval top-1/top-5, coherence distributions and a random baseline for item-by-item checking.
Four commands, verify it yourself
No upload, no sign-up, no GPU required.No upload, no sign-up, no GPU required.
git clone https://github.com/zhibaiYingChuan/DaoTi.git
cd DaoTi
pip install -r requirements.txt
python scripts/eval_benchmark.pyThe benchmark script depends on no unpublished component. Output goes to benchmarks/benchmark_results.json. If you do not want to set up an environment, run the Colab notebook in the browser — full inference takes about 5 minutes on a free CPU.
| Item | Min / recommended | Min | Recommended |
|---|---|---|---|
| Memory | 2 GB → 8 GB | 2 GB | 8 GB |
| Storage | 50 MB free → SSD | 50 MB free | SSD |
| CPU | x86_64 with AVX2 → 4 cores at 2.5 GHz or better | x86_64 with AVX2 | 4 cores at 2.5 GHz or better |
| Python | 3.8+ → 3.10+ | 3.8+ | 3.10+ |
| PyTorch | 1.13+ → 2.0+ | 1.13+ | 2.0+ |
GPU inference is optional, roughly 3–5× faster, with a 1 GB minimum of VRAM.
Fifty-one versions of the road here
From V50 to V91, every step has an experiment log.From V50 to V91, every step has an experiment log.
- Foundation · V50 — V53
- Four-stage curriculum:启蒙 → 筑基 → 通识 → 专业.Top-1 93.8% → 96.1%
- Emergence · V54 — V57
- Key finding: freezing pretrained weights beats full fine-tuning. The paradigm is established.Top-1 98.0%
- Insight · V60 — V70
- Cross-domain isomorphism confirmed on text_proj; multi-config generality verified.Dao-isomorphism early stop at epoch 77
- Proof · V80 — V91
- The degenerate ground state is written as a formula; multiple equivalent solutions observed directly.CKA 0.27 / top-1 100%
Training data is classical Chinese corpora with no AI-generated content; all training ran on consumer CPUs. The full experiment chain is in chapter 4 and appendix A of the whitepaper.
Draw the structure
Three interactive concept visualisations. They show structure; they do not run the model.Three interactive concept visualisations. They show structure; they do not run the model.
All three are structural visualisations; the numbers on screen are illustrative approximations, not real inference output. To see real output, run the model locally or in Colab as described in the previous chapter.
What you can and cannot do
Spelling it out is more useful than saying “open source”.Spelling it out is more useful than saying “open source”.
- Code · Apache 2.0
- Inference, adapter training, benchmark and demo code may be used commercially and modified, provided you keep the copyright notice and the patent grant.
- Weights · Research License v1.0
- Inference use is permitted. Three things are explicitly forbidden: reverse engineering the weights to extract the training method; training or distilling a competing model from them; redistributing or reselling the weights as a standalone product.
- Papers
- All six are working documents by an independent researcher and have not been peer reviewed. Judge them yourself before citing.
- Terminology
- “Gauge field”, “degenerate ground state” and “Goldstone mode” are engineering analogies. We do not claim to have discovered a law of physics.
- Security boundary
- Domain gating and coherence constraints stop specific input types, but do not guarantee coverage of every attack form.
Where to start
Read the code, run the benchmark, ask a question, or just write to us.Read the code, run the benchmark, ask a question, or just write to us.
Architecture source, training code and data recipes are not in the public repository and require separate authorisation.
spring60@foxmail.com