Skip to content

Chapter 8 — Operations

AI Improvement Pipelines

Fine-tuning pipelines integrated into the Conversation Monitor, covering speech-to-text models, Large Language Models and the text-to-speech vocabulary. Each pipeline runs three stages: data preprocessing, fine-tuning and evaluation.

1. Data preprocessing

Preprocessing aggregates all improvement data collected in the platform within a given period and converts it into the format required by the specific model. When the process finishes, the output is cached and stored for the fine-tuning phase.

Figure 7 — Data preprocessing step

Figure 7 — Data preprocessing step.

2. Model fine-tuning

To create a fine-tuning task, open the Models section and click Create new version . The form defines the metadata and configuration required to start the pipeline:

  • Name and Description — required fields for the new model version.
  • Model and Model architecture — the model to fine-tune and its associated architecture.
  • Parent model version — an existing version used as the basis for the new one.
  • Use language model — indicates that the uploaded model is a language model.
  • Upload Model and Upload Processor — the model file and the processor (tokeniser) file.
  • Create — starts the fine-tuning pipeline once all information and files are provided.

The dialog displays the current upload status and prompts for both files before proceeding.

Infrastructure requirement

GPU-enabled infrastructure must be connected to the platform. Without the required computational resources the training pipeline cannot start.

Figure 8 — AI model fine-tuning form

Figure 8 — AI model fine-tuning.

3. Evaluation

Completed model versions are evaluated against predefined datasets to produce measurable performance indicators for validation and comparison. The evaluation interface presents a tabular overview of all evaluation jobs, one row per instance:

Field Description
Id Identifier of the evaluation record.
Name Name assigned to the evaluation.
Description Descriptive text associated with the evaluation.
Confidence score Confidence level calculated for the evaluation.
WER Word Error Rate.
CER Character Error Rate.
Dataset Dataset used in the evaluation, such as Validation or Test.
Total Summary indicator of the evaluation task.
Progress Percentage and progress bar showing the execution state.
Status Execution status of the evaluation, such as Finished.
Delete / Open Delete the evaluation entry, or open its details.

The Create Evaluation button in the upper right starts a new evaluation. Once results confirm the model meets the required quality criteria, it can be published and made available in BlockFlow — so only validated and approved versions are exposed for production use.

Figure 9 — Evaluation window

Figure 9 — Evaluation window.