yanqiangmiffy/chinese-decision
User repository on GitHub
- Forks
- 7
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Models & Reimplementations
Chinese Decision
A Chinese-language decision model that returns a probability for each fixed candidate option in one forward pass, with no generated text or chain of thought. It is a LoRA fine-tune of Qwen3.5-0.8B on 13,048,000 CJ-Dataset rows, benchmarked on CJ-Bench against Chinese-Jev.
- Stars
- 27
- Language
- Python
- License
- MIT
- Reading time
- 2 min
- Published by
- Community
- Author
- yanqiangmiffy
- Last verified
- Oct 2026
Highlights
- One forward pass returns a probability per option from the candidate-letter logits A-Z; no autoregressive generation and no chain of thought.
- Three question types: noul (yes/no), choice (single choice) and score (ordered levels returning an expected value).
- README comparison on CJ-Bench: 72.24% general-domain accuracy, above closed Jev at 68.35% and Chinese-Jev General at 69.20%.
- Calibrated without temperature scaling: overall ECE 1.07% on CJ-Bench, 2.46% average across medical, legal and finance; medical accuracy 88.86%.
- 0.8B parameters, about 70 ms median per item on one A800 and 6.7 GB peak VRAM; a Q8 GGUF build runs on CPU, and MIT covers the release.
Watch out
README reviewed 2026-10-17; it is in Chinese and the digest truncates the news section mid-sentence, so the training and release story is incomplete. All benchmark numbers — CJ-Bench accuracy, ECE, latency — are the project's own comparisons against Chinese-Jev and closed Jev, not independent evaluations.
Questions about this entry
Short answers drawn from the linked project or page.
A lightweight Chinese-language decision model from yanqiangmiffy: given a context and a fixed set of candidate options it returns a probability for each option, and it does not chat or generate text.
It reads the logits for the candidate letters A through Z at the last position and applies softmax, so one forward pass yields the distribution; there is no autoregressive decoding and no chain of thought.
Three types: noul for yes/no, choice for single choice and score for ordered levels with an expected value. The README lists intent recognition, ticket routing, moderation, labeling and agent step selection as fits.
On CJ-Bench the README reports 72.24% general-domain accuracy versus 68.35% for closed Jev and 69.20% for Chinese-Jev General, and a 76.98% professional-domain average against Jev's 77.01% — the project's own comparisons.
It is a LoRA fine-tune of Qwen3.5-0.8B on 13,048,000 Chinese decision rows (CJ-Dataset): one epoch, 10,132 steps, with the best validation step 10000, using a method adapted from n4ze3m/hmm.
The full model runs on a single A800 with about 70 ms median latency per item, roughly 132 items per second in batch and 6.7 GB peak VRAM. A Q8 GGUF build runs on CPU, and the companion Decision Studio starts with docker compose up.
The digest truncates the news section, so the training and release story is incomplete, and no hosted endpoint or pricing is documented. Accuracy, ECE and latency figures are self-reported against Chinese-Jev and closed Jev, not reproduced independently.
More Jev resources like this
An open-source, local alternative to Jev: System-One decision models that answer the POST /v1/systemone shape — choice, yes/no, or score — for general decisions and browser-agent steps. Models ship as OpenJev-1.7B, 4B, and 8B, with a pip-installable Python package.
Jev Models & Reimplementations#community#github#python
Community
alanhuangyoo/openjevPython
alanhuangyoo/openjevPythonAn open System One decision model from Interfaze-ai: give lev a state — text, a ticket, an email, or JSON — plus yes/no, choice, or score questions and it answers all of them in one forward pass from logits, returning calibrated probabilities and zero output tokens.
Jev Models & Reimplementations#community#github#python
Community
abhinavexists/levPython
abhinavexists/levPythonThree open-weights 4B decision models for English (Deem-4B), Turkish (Karar-4B) and German (Wähler-4B) that choose between options and return a probability for each. The server implements the Jev API, POST /v1/systemone, so existing Jev clients work by changing the endpoint.
Jev Models & Reimplementations#community#github#python
Community
mertkayacs/jevaltPython
mertkayacs/jevaltPythonA small, fast decision model — a system one model — for agentic workflows from Strands Labs: it picks between options or rates on a scale faster than an LLM and returns a calibrated confidence with every decision.
Jev Models & Reimplementations#community#github#python
Community
strands-labs/strands-deciderPython
strands-labs/strands-deciderPythonSmall Jev-style decision models that read a state and typed questions and return a probability for every option in one forward pass without generating text; the 0.75B and 4.2B versions speak TypeSafe's Jev API format at a median of about 50 ms per request on one H100.
Jev Models & Reimplementations#community#github#decision-model
Community
jiwidi/jiwoPython
jiwidi/jiwoPythonSystem One models from first principles by Qi Li and Xinchao Wang at the National University of Singapore: give Jeva a state and typed questions and it returns candidate probabilities, a Boolean probability, or an expected rubric score, with reusable option memory.
Jev Models & Reimplementations#community#github#python
Community
liqiiiii/jevaPython
liqiiiii/jevaPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.