Prepare a sensor handover from a note containing a similar sensor ID, a hypothetical report, a numeric reading and a negated flag.
Laya / MaintenanceResearch preview
Maintenance Decision Assist
Read reported sensor flags; explore maintenance decisions as experiments.
Start with the fine-tuned task, or explore an experimental decision below. Experimental fine-tune: results can be wrong; confidence is not accuracy.
This model was fine-tuned to extract the latest explicitly reported sensor flag. These fictional examples were used in training. The other tabs explore tasks not trained here.
Fine-tuned task · Choice
What is the latest reported flag for this sensor?
Read the note, run the model, then compare its answer with the training reference.
Same synthetic test: fine-tune 57.4%; exact-rule parser 100% for the tested grammar. The model has not shown an advantage over those rules.
Choose a training example
A log with distracting entries
Loaded original: synthetic training example · Edits are new inputs
Where this could fit: preparing a handover from written notes. Structured logs can be handled by rules; this model has not demonstrated an advantage over them. This output reports a flag, not equipment health.
Try a change
Change the actual 08:08 report for S6-101 to FAULT.
Compare the model’s answers
Run the example to compare the possible answers.
Live model result · Confidence included
Exact input and response for this run. Editing or resetting clears the pair.
Source: data/semantic-v3/admitted/maintenance-targeted-train.json · row targeted-slot128-o1 · source case v3-0fb13088ea75cecf. The loaded original state and question were seen by selected update 64. Edited input is no longer the original training example. Synthetic wording may say ‘actual’; those events are fictional.
Decision needed · Experimental taskChoice
Which machine should get the next available inspection slot?
One technician can inspect only one machine before the next shift. Decide which job to put first using the supplied policy.
Experimental workflow · Fictional example · Not a task trained in this fine-tune
1 · Review the inputs
Read both inspection requests and the site policy. Choice selects Pump A, Pump B, or asking the shift lead for more information.
Try a change
Switch which machine has the repeated alarm.
2 · Read the model result
Compare the model’s answers
Run the example to compare the possible answers.
Live model result · Confidence included
3 · Your next step
The shift lead reviews the suggested priority and assigns the inspection slot. This is not a stop-or-run decision.
Run an example to see the input and output for that run. Editing or resetting clears the previous pair.
Decision needed · Experimental taskScore
Use this maintenance window, or look for a less disruptive one?
The maintenance has not started. Estimate how the proposed stop would affect output before choosing a window.
Experimental workflow · Fictional example · Not a task trained in this fine-tune
1 · Review the inputs
Read the proposed work, stock buffer and backup capacity. Score rates the expected output disruption on the three levels below.
0The plan preserves normal output through available stock or alternative capacity.
1The plan preserves some output, but at reduced capacity.
2The plan interrupts all affected output during the maintenance window.
Try a change
Compare with a planned stop fully covered by finished stock.
2 · Read the model result
See where the evidence lands
Run the example to place it on the defined scale.
Live model result · Confidence included
3 · Your next step
Compare the estimated output impact with other available windows. A planner decides the timing; this result does not establish whether delaying work is safe.
Run an example to see the input and output for that run. Editing or resetting clears the previous pair.
Decision needed · Experimental taskNoul
Book a technician slot, or resolve missing planning details first?
For job P-17, the planning checklist requires a reserved seal kit, an assigned qualified technician and an agreed production window. Review these before booking.
Experimental workflow · Fictional example · Not a task trained in this fine-tune
1 · Review the inputs
Statement to assess: “All three planning prerequisites are documented.” Read the job note for the part reservation, technician assignment and agreed window.
Try a change
Compare with an unreserved part and a tentative window.
2 · Read the model result
Explore one clear statement
Run the example to see the model’s true / false probabilities.
Live model result · Confidence included
3 · Your next step
Confirm each prerequisite in the source note. If any is missing, resolve it before booking; if all are documented, the planner can review the booking. The model result is not work authorization.
Run an example to see the input and output for that run. Editing or resetting clears the previous pair.
Try your own decision
Choose the Laya output type first. The form will change to collect exactly what that type needs.
Choice: your possible actions → one selected action. Score: your ordered levels → a position on that scale. Noul: your statement → its true/false probability.
Next: service triage. Explore what information to request and when to route an equipment issue to a technician or specialist.
Later: warranty and service-claim review. Help identify missing evidence and route claims to a reviewer.
These are planned directions, not capabilities of this checkpoint. They need suitable data and separate validation; no release dates are promised.
If the queue is full, the connection drops, or controls stay disabled, reload to start again. This clears your inputs and results; copy any notes you want to keep first.
The model extracts explicitly reported sensor flags. It does not diagnose a fault or establish whether equipment is safe. Selected v3 accuracy: 57.4% on the small synthetic study.
On the same 64-case synthetic test: uniform four-option guessing 25%; base Laya 39.8%; this fine-tune 57.4%; an exact-rule parser 100% (64/64), limited to the tested grammar. Older synthetic task results: Score 25.0%, Noul 48.5% accuracy. These older tasks do not validate the experimental demos.
These workflows explore interpreting written evidence. We have not established an advantage over rules; structured fields and fixed numerical policies are usually better handled directly.
V3 targeted / fixed seed 7601. This model did not meet every study progress safeguard. 64 synthetic cases in 16 groups; four option orders are correlated. No expert or real-world validation. Training is frozen.
The app does not intentionally save your notes, call another language model or take actions. Use non-sensitive examples. On Hugging Face, inputs are sent to the Space for inference; hosting infrastructure may maintain operational logs.
Base model: Convai Innovations Laya. Independent fine-tune; no upstream or TypeSafe endorsement. Model cards include the complete evaluation and provenance.