Task and data design
Define source-grounded answers, clarification and authorized reads as concrete behaviors. Separate current business facts from information that belongs in shared weights.
Intueo Labs / Language model research
We’re developing language models for grounded answers, clear requests and carefully scoped business workflows. Built through task-specific post-training and evaluation, with Intueo Platform as the intended working environment.
Not generally available. A release version has not yet been assigned.
Conceptual illustration, not a model architecture or product screen.
01 / The problem we are working on
Business work is full of details a plausible answer can get wrong: the revision of a document, the account a record belongs to, or whether a request is asking for information or permission to act.
IntueoLM studies how smaller language models can handle those distinctions. The research target is teams reviewing documents and authorized business activity, with English and French examples grounded in the workflows we build around Intueo Platform.
To investigate task-specific behavior at the model layer, not only adjust prompts around a general assistant. We define the task, establish a baseline, train a bounded adapter experiment and inspect whether it improves the behavior that matters. A branded checkpoint is not the goal; a demonstrably useful improvement is.
Illustrative, fictional scenarios. These are target behaviors, not generated IntueoLM outputs or a live demo.
What is the inspection window for Line A?
Use the current revision, retain the exact line and time, and identify the source. Do not fill gaps from a plausible but outdated schedule.
02 / How we are building
IntueoLM is a model-development programme, not a renamed inference API. We have trained real adapter weights on an existing open-weight base. This is not foundation-model pretraining from scratch.
Define source-grounded answers, clarification and authorized reads as concrete behaviors. Separate current business facts from information that belongs in shared weights.
Train a small set of model parameters while retaining the underlying base. Preserve the exact recipe, tokenizer and weight identity so the experiment can be inspected.
Compare with the same base under matched conditions. Inspect failures and distinguish a model proposal from permission-controlled execution in the application.
Experimental technical profile
The documented pilots, not final IntueoLM product specifications. Base selection remains open.
The experimental base is Apache 2.0. That does not assign a licence to Intueo’s adapters or datasets. The research note records the pinned revision, data conditions and retained artifact identities.
851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0aRead the technical note03 / What the experiments show
Two internal pilots established training and artifact-reload mechanics. Neither met the bar for deployment. We show the paired results and the decision together, including the regressions.
Internal diagnostics, not a public benchmark
The first adapter omitted source identifiers in four grounded answers. Its lower diagnostic count and inconsistent clarification behavior did not establish a useful improvement.
Passed checks / total checks on small fictional fixtures. Compare base and adapter within a pilot only: cases, contracts and output caps differ between pilots. These are not safety accuracy, general benchmark scores or independent human evaluations.
The checks include lexical markers and native-format selections. Some first-pilot citation expectations were not explicit in every input contract. The test outputs have now been inspected and cannot be reused as a fresh blind test for future tuning.
04 / Data and reproducibility
The pilot training data and the current design corpus are different things. The pilots used small selections from offline, AI-assisted fictional templates. Our next design work focuses on read-only Platform workflows, exact source context and reviewable labels.
Five simulated read proposals, six tool catalogs and 18 assistant payloads. Actual formatting, response-masking and CPU-collation helpers preserve the intended content. No model was loaded for that check.
Design and data-path evidence, not approved training data or model performance. Token counts describe these examples, not a supported context window.
The documented experiments used no customer conversations or private agent memory. No real business tools were executed.
No teacher API calls were used in the pilots. AI-assisted authorship still needs applicable output-use review, and no external corpus has yet been cleared for ingestion.
Domain and French labels, missing provider schemas and representative untouched tests need review before a new training run. Automated formatting checks do not replace it.
A BF16-LoRA reference forward/backward check at batch 1 and 543 tokens produced finite gradients in all 496 adapter matrices, with zero optimizer updates and unchanged parameters. Peak PyTorch allocation was about 9.5 GiB. This is not a full-training or serving-memory guarantee; six current design examples are longer than that tested input.
05 / The model and the product
Language-model behavior, persistent context, execution permissions and the customer interface solve different problems. IntueoLM is intended to fit into that system, not replace the checks around it.
The research model interprets the supplied input. Its output may propose a read, ask a question or answer from evidence. A proposal is not execution.
06 / Research resources
Start with the technical note and machine-readable research snapshot. They document the current approach, measured observations, limitations and the next release gates.
Purpose, experimental method, data, paired results and lessons from the pilots.
Read the noteStructured observations, exact counts, artifact identities and retained report hashes.
Inspect the dataGitHub research materials and a future Hugging Face model card are being prepared. Public repository and model links will appear here when an approved package is actually available. No public weights, model endpoint or open-source licence is announced by this page.
No. Current work is adapter-based post-training of an existing open-weight base. Intueo develops the task design, experiments, evaluation and integration approach. Final base selection remains open.
Not currently. Both documented trained candidates were rejected for deployment. There is no accepted public model release, downloadable weight package or IntueoLM serving endpoint.
Those product specifications have not been established. The token counts here measure draft inputs, and the memory figure describes a single reference gradient check. Neither is a production specification.
No such claim is being made. Platform is a separate working product with configured agents and models. A future IntueoLM integration requires its own evaluation and approval.
Useful conversations start with a task, sources, permitted actions and a clear way to judge the result. We welcome discussion around grounded business workflows and English/French evaluation.