chaoliangUNSW/MacJev-322M-4K-Laya

MacJev-322M-4K-Laya

A compact decision model for local Mac agents: one forward pass, calibrated probabilities, 4096-token inputs.

MacJev takes an observed state and a question with candidate answers. In a single non-autoregressive forward pass it returns a probability for every candidate. Use it to rank candidate actions, route tool calls and check task state, in Chinese or English, entirely on your machine.

[image]

Highlights

Compared with the Laya multilingual checkpoint it was trained from, with the same 4096-token input budget, on held-out test sets:

• —Reads long observations. Yes/no task-state checks on 2K–4K-token inputs rise from 31% to 89% accuracy.
• —Probabilities you can threshold. Calibration error on long inputs is about 8 times lower, so stated confidence closely tracks actual accuracy.
• —Better on public benchmarks too. Across 11,600 public decisions from typed decisions, Emotion and AG News, accuracy rises by 1.5 points, with a 95% interval of 1.1 to 1.8.
• —Reproducible. The long-input and public-benchmark gains appear in all three independently trained seeds. This release is the seed chosen in advance on development data.

Why MacJev for agents

• —One pass, every candidate. All options are scored together in one forward pass, with no token generation.
• —Your answer set, per request. Choice, ordered Score and yes/no Noul questions take options you define at request time. No classifier head to retrain.
• —4× more room. 4096 tokens of total input and 1024 for the question and options, four times the defaults of the checkpoint it starts from.
• —Nothing silently dropped. Inputs over budget raise an error instead of quietly cutting candidates or evidence.
• —Calibrated per question type. One temperature each for Choice, Score and Noul, fitted on 6,000 independent calibration decisions.
• —Trained for the Mac. 20,000 Mac decisions built from verified trajectories across files, Chrome, Word, Excel and PowerPoint, including recovery cases, alongside general-language, typed and rule decisions.
• —Local and decision-only. No API calls. MacJev returns decisions and never executes actions, so your agent keeps its own executor and confirmation step.

Formats

MacJev is a decision model rather than a chat model, so each format ships with a small runtime that runs the full model, including its decision head. To use MacJev from a chat assistant such as LM Studio, expose decide() as a tool or MCP server.

Quickstart

hf download chaoliangUNSW/MacJev-322M-4K-Laya --local-dir MacJev
cd MacJev
python -m pip install -r requirements.txt

from macjev import MacJev

model = MacJev(".", device="mps")  # Apple silicon; use "cpu" or "cuda" elsewhere
result = model.decide(
    state={
        "request": "打开浏览器",
        "available_actions": ["open_browser", "copy_file", "ask_user"],
    },
    question={
        "t": "choice",
        "ins": "Which available action best matches the user's request?",
        "crit": {
            "open_browser": "Open the user's browser",
            "copy_file": "Copy a file inside the approved folder",
            "ask_user": "Ask for clarification",
        },
    },
)
print(result["answer"], result["probabilities"])  # A decision only. Nothing is executed.

The result contains answer, probabilities, top_probability, input_tokens, latency_ms, and actions_executed: False. Keep one model object loaded across requests.

Other question shapes:

score = {"t": "score", "ins": "How complete is the observed task?",
         "crit": ["Not started", "Partly complete", "Complete and verified"]}
boolean = {"t": "noul", "ins": "Does the observation prove that the requested file exists?"}

Score answers are ordered string indices ("0", "1", ...). Noul answers are "false" and "true". Choice keys keep their insertion order.

The CLI reads one JSON object per line with state and question:

python macjev.py --model . --device mps

Training

• —Base: convaiinnovations/laya, multilingual subfolder, revision 1c5edc17a7acd8701df6fc341c0d179f1c62c982. Encoder lineage jhu-clsp/mmBERT-base.
• —Size: 321,908,995 parameters; 44,861,185 adapted. The last six encoder layers, type embeddings, decision transformer and scorer were trained, starting from the official decision head.
• —Data: 50,000 decisions split evenly between short and long inputs: 25,000 up to 1K tokens, 12,500 from 1K to 2K and 12,500 from 2K to 4K. A separate 3,000 development and 6,000 calibration decisions.
• —Objective: supervised fine-tuning with an ordinal loss, replay of verified original-model outputs, and candidate-permutation consistency, trained in stages from 1K to 4K inputs.
• —Selection: recipes were compared on development data, then three seeds were trained from the official weights. Seed 0 was designated in advance; test results were never used to choose it.
• —Hardware: one NVIDIA H100, FP16 encoder arithmetic with an FP32 decision head and master weights. Released weights are FP32.

manifest.json lists every file's checksum, and the loader verifies them. validation.json records the loading and integrity checks.

License and attribution

Apache-2.0, inherited from Laya; see LICENSE and NOTICE. The mmBERT-base backbone is MIT-licensed; see MMBERT_LICENSE. MacJev is an independent project. It is not affiliated with or endorsed by JEV or the Laya authors, and it contains no JEV weights.

• —Laya model and source
• —mmBERT-base

Community

Interested in compact decision models, local agents, and practical macOS automation? Join our Discord community.

如果你也对轻量决策模型、本地 Agent 和 macOS 自动化感兴趣，欢迎加入我们的 Discord 社区。