A LoRA adapter needs the right base model
Review the adapter together with its base model, tokenizer and configuration. Separate the file, loading, activation and the quality of the setup actually run.
Article language: English

In this article
The package contains an adapter, the label “version for review” and a link to a family of models. It does not identify the exact base model. At a fictional company preparing exhibit descriptions, this small omission is enough for two people to assess two different setups even though they both opened a file with the same name.
LoRA is a way to fine-tune a model using an additional set of parameters. An adapter delivered separately is not a complete standalone model. Reviewers need to establish what it was attached to, how the input was prepared and whether it actually contributed to the answer being assessed. A label saying “our model” fits easily on a folder. The rest of the setup may not.
A small file does not describe the whole release
The PEFT checkpoint documentation explains that an adapter checkpoint contains its parameters and that a base model is also needed to run it. It describes a separate configuration file that can contain information about the base and revision. The presence of such a field alone does not mean it identifies the exact state used in the assessment.
In our fictional company, the adapter is meant to help preserve the format of descriptions: the exhibit name, its origin and a separate note on missing information. The team is reviewing a change in behaviour, not just the ability to download a file. A base family name is too broad if it covers different releases or variants. Distinguish the base identifier from its particular revision.
A release manifest is simply an inventory of the components needed to reproduce the setup. It should identify the base, adapter and configuration files and their origins. A checksum helps reveal a change in content. Recording the runtime version shows what read those files. This is a document for the provider and the person accepting the work, not an assignment for every description writer to memorise parameters.
Separate “base X was expected” from “base X at revision Y was loaded”. The first is a release condition; the second is an observation from a particular run. A report that shows only an intention recorded in configuration does not establish the actual state. The difference should be clear without comparing several unlabelled screenshots.
The adapter also needs its own version. A folder called “latest” can change while a copy on another computer remains older. With each report, retain the identifier of the file actually used. Do not attribute a later corrected adapter to answers recorded before it was enabled.
An incomplete manifest stops the comparison before anyone asks about the style of the descriptions. If two base files have the same size but different contents, size will not identify the right one. If the revision is missing, the provider should establish it from the retained preparation records rather than enter today's default version. The report can honestly retain the limit: “the base used in the previous assessment was not reproduced”. That is more useful than an apparently complete inventory filled in from memory.
If the quantization method also changed (in Polish), the representation of the weights is another property to record in the manifest. An earlier assessment does not automatically carry over to the new combination. There is no need to assume it will be worse. You do need to know which setup is being reviewed now.
Loaded does not always mean active
A clean start is not yet evidence that an adapter was used in an answer. An application may run the base alone, select another adapter or route the task to a fallback executor. These are possibilities to identify in the particular product, not a claim about every library that supports LoRA.
The PeftModel reference distinguishes available, active and merged adapters and states that a newly loaded adapter is not automatically made active by load_adapter. Its status tools help distinguish those states. This is behaviour of a specific programming interface; other loading methods require checking their own documentation.
A technical specialist should provide evidence of state linked to the run being assessed. A startup log from last week does not prove activity during today's trial. Where several adapters are available, selecting the right set for the task also matters. A list of available names alone does not answer which one was used.
A business reviewer does not need to interpret every model layer personally. They can require a clear answer: which variant ran the test, on which base and with which adapter? If the technical status is inconsistent or cannot be associated with an answer, the quality result remains insufficiently described for a release decision.
Asking the model “are you using the adapter?” does not replace that evidence. Its statement is a language response, not an independent measurement of its configuration. The principle that an AI explanation is not an execution trace (in Polish) applies even to a polished assurance about its own version. Confirmation must come from the appropriate part of the runtime process.
An Aeneia Audit operational review can organise these requirements, responsibility for evidence and the link between a test and a release. The review does not replace a technical compatibility analysis of every architecture. It can help a team avoid approving behaviour without knowing what generated it.
The text in the window is not yet the model's input
The setup also includes a tokenizer, the mechanism that converts text into units used by the model. Its files and settings should match the agreed configuration. If the vocabulary was changed or special markers added during fine-tuning, the provider should clearly show how that state was reproduced for the review.

The chat template is a separate component. It defines how messages and participants' roles are presented to the model. The Transformers documentation shows different control tokens even for variants derived from the same base and explains why the appropriate conversation format matters. Similar-looking chat windows therefore do not establish identical input.
In the exhibit-description example, a system instruction specifies the output fields and the user's message contains material about the object. Preserve those roles and how they are passed to the model when reviewing the setup. If a new application combines everything into one text or adds an instruction of its own, it changes the test conditions. Comparing only the sentence entered by the author is not enough.
If a special marker for missing data was added during fine-tuning, include its expected meaning in the test. The company may require it in a separate field rather than in the finished description for visitors. Check the final display after the application has processed the answer too. A correct marker in the raw output is not sufficient if a later step removes it or shows it to the audience as an unintelligible string. This checks the transition from the model setup to a usable description.
Also record length limits and how the answer ended. A missing final field may result from generation being interrupted rather than from the fine-tuned variant omitting a rule. Where the system provides it, the report should retain the reason for termination and the actual output. A truncated description should not be assessed as a complete answer just because its opening has the right style.
For the first comparison, prepare the material in the same way for the base and the adapter-equipped variant. If you intend to deploy a different template, assess that as an additional change. Conclusions can concern the whole solution, but they should not attribute every difference to the adapter file alone.
A control test should reveal a mixed-up setup
First confirm that the required combination is available and can be reproduced in the agreed environment. Then test its behaviour. The order matters: without identifying the configuration, even a good result cannot be assigned to the right setup. Conversely, a correct manifest does not establish that the answers are useful.
In the fictional case, the exhibit material has a name and place of origin but no date. The required format must identify that missing information separately. This is precisely the division the adapter was intended to support. The example input describes “a ceramic vessel from a workshop in Toruń” and gives no year of creation. A correct result retains the name and origin and notes that dating is missing. Adding “19th century” is an error regardless of the elegance of the remaining text. Omitting the separate missing-information field is an error in the agreed format. The review record can mark these two error types separately. The reviewer receives the input and the required fields so they can assess the handling of data and format rather than a general impression of museum style.
A test with the base alone and one with the active adapter can help reveal a difference, provided the other conditions are kept the same. No difference in one easy case does not prove the adapter is inactive. The base may already meet the criterion. A difference is not sufficient evidence of activation either: it may come from different settings or generation variability. Technical observation and task assessment therefore remain separate evidence.
Fine-tuning a model and preparing its training set (in Polish) involves choosing behaviour to repeat. Here we check the final setup rather than reproduce training. Review cases should differ from demonstrations chosen during preparation. An answer remembered from a demonstration does not settle how the system will handle new descriptions.
Add a controlled case with a missing required adapter if it can be run in an isolated environment. The purpose is to check how the application responds, not to experiment with client material. Does it stop the task with a clear message or fall back to the base alone? If fallback is permitted, the result should label it clearly. An adapter cannot be accepted on answers produced without it.
Once the complete setup is restored, close the test environment and recreate it from the delivered package using the provider's instructions. The aim is to uncover reliance on a file left behind by chance or on a previously active adapter. In the new run, retain the state and answer again. If reproduction requires a file from the author's computer, the release is not yet complete within the agreed scope. This test does not replace checks of permissions or weight provenance; it answers the narrower question of whether the setup can be reproduced.
What is the reviewer actually signing off?
In the review record, separate completeness of materials, confirmation of the configuration that ran and the task assessment result. A shared green “works” field is not enough. Each can need a different correction. A missing base version calls for release documentation; an incorrect answer format requires investigation of the setup's behaviour.
For a merged release, where the adapter's effect has been incorporated into the base weights, do not automatically expect a separate adapter switch. The complete file needs provenance documentation and an assessment of that form of release. The PEFT checkpoint reference describes merging and its limitations; not every method and configuration supports it. The delivery record should match the artifact actually supplied.
Keep the previously approved setup and the conditions for returning to it. During a later model migration (in Polish), the base may change independently of the adapter. Do not assume compatibility from a similar name alone. A new combination needs its own scope of checks and an explicit decision.
If the solution must run without a network, assessment of the local AI workflow (in Polish) includes availability of every required file. A small adapter without a downloaded base will not perform the task in a room without a connection. A quality report and an availability test answer different questions; do not replace one with the other.
Acceptance should therefore refer to an identifiable setup and a recorded test. For the next exhibit description, the team needs to know which release it is using and where the confirmation obtained ends. The adapter name can remain short. The evidence of what actually ran should be more precise.







