Taiji-2B 太极

Fast intuition and slow reasoning, fused into one agent system. An open System One model that decides and writes in a single 2B model, and hands the hard cases to any System Two LLM, the way yin and yang form one circle.

Qwen3.5-2B with a LoRA (rank 16) and a decision head. Give it a state (text, JSON, a screenshot) and questions with the options you allow; it returns calibrated probabilities over exactly those options in one forward pass, without generating answer tokens. The same weights also write short text, such as a form-field value or a tool argument. When it is unsure, route the question to any larger LLM.

Code, server and examples: github.com/nigelleong0703/Taiji. The request format is compatible with the /v1/systemone choice / noul / score questions of TypeSafe's Jev (independent project, not affiliated).

What it does

• Decide: choice (1-255 options), noul (true/false), score (2-10 ordered levels), all options scored in the same pass; the answer is always one of the offered options.
• Write: the text for one field or argument, or null when the context lacks the value, from the same weights.
• Escalate: temperature-calibrated probabilities (held-out ECE 0.02) tell when to ask System Two.
• See: an optional screenshot goes through Qwen3.5's vision encoder next to the text state.

Results

General decisions vs hosted Jev, same 490 held-out questions, same script and grading:

Questions	n	Jev (hosted)	Taiji-2B
All	490	0.859	0.882
General choice	248	0.891	0.899
General yes/no	33	0.970	0.970
General score	19	0.632	0.684
Tool choice	116	0.871	0.914
Tool yes/no	74	0.743	0.784

On par (the models disagree on 57 questions: 34 vs 23, McNemar p ≈ 0.19). The questions come from the same public datasets as part of the training data (held out), which favors this model.

Held-out decisions (2,945 rows): accuracy 0.86, ECE 0.02. Field text exact match 0.76. Operation recall: BLOCKED 1.00, WAIT 1.00, SELECT 0.96, CLICK 0.93, TYPE_TEXT 0.89, DONE 0.57, SCROLL_DOWN 0.42, SCROLL_UP 0.18 (calibration_report.json).

Speed: 219 ms median server time per decision on a real Google Flights page (~6,500 tokens with a screenshot and three questions) on one H100 (171 ms model, 59 ms input preparation); network time comes on top. About 6 GB of VRAM.

Browser agent (one example integration): simple sites 6/6 with a correct completion; Google Flights reaches the right results page with a screenshot but does not yet declare completion. Browsing in this page format is under-represented in the training data.

Use

git clone https://github.com/nigelleong0703/Taiji && cd Taiji && pip install -r requirements.txt
hf download nigelleong0703/Taiji-2B --local-dir Taiji-2B
pip install --no-deps Taiji-2B/wheels/causal_conv1d-*.whl   # optional kernel: torch 2.14 + CUDA 12.6
python infer.py --adapter Taiji-2B --data examples/requests.jsonl --output answers.jsonl
S1_API_KEY=<secret> python serve.py --base Qwen/Qwen3.5-2B --adapter Taiji-2B --shared-prefix

This is the project's adapter + head.pt format, loaded by s1.load(); not an AutoModel.from_pretrained repository.

Files

File	What
adapter_model.safetensors, adapter_config.json	LoRA rank 16 on the Qwen3.5-2B language model
head.pt	decision head (yesno: Yes-minus-No readout at each option marker + gated candidate head)
temperature.json	calibrated temperatures, global and per question kind
calibration_report.json	held-out report per source and kind, operation recall, DONE analysis, text exact match
train_args.json	training arguments
wheels/	causal-conv1d 1.7.0 built for torch 2.14 + CUDA 12.6, sm 80/86/89/90, Python 3.10

Training

One epoch over 60,395 rows on one H100 (~5.4 h): LoRA rank 16, vision encoder frozen, cross-entropy + 0.1 Brier, label smoothing 0.05, text rows weighted 3x, then temperature calibration. Data: WebLINX 12,000 and Multimodal-Mind2Web 9,000 web steps, Open-Jev browser 13,814 and general 10,000 decisions, tool selection 3,000, general public decision sets 3,000, synthetic WAIT/BLOCKED 1,500 and SELECT 2,952, completion rows from WebLINX 1,420.

License

CC BY-NC-SA 4.0, research and non-commercial use only: the training data includes WebLINX (CC BY-NC-SA 4.0). Built on Qwen3.5-2B (Apache-2.0); other upstream datasets keep their own licenses. The code on GitHub is Apache-2.0. A DONE or any other choice is not proof of success: verify outcomes independently.