sft_prompts

Prompt and chain-of-thought templates for the three tasks, plus the template-filling helper used to instantiate them.

Prompt templates and helpers for MedVision SFT and benchmark evaluation.

This module collects the task prompts, format requirements, and Chain-of-Thought (CoT) reasoning templates shared by supervised finetuning (SFT) data formatting and MedVision benchmark evaluation. Because these strings are consumed by both pipelines, editing a prompt here changes both training targets and evaluation behaviour, so modify with care.

Two prompt families are provided: plain format prompts used in non-CoT benchmarking, and CoT instruction/template pairs (for tumor-lesion size, distance, angle, and detection tasks) whose placeholder tokens are filled in by fill_in_template().

medvision_bm.sft.sft_prompts.fill_in_template(template, values_dict)

Substitute placeholder tokens in a CoT template with concrete values.

Iterates over values_dict and replaces every occurrence of each key in template with str(value). Used to fill the <...> placeholders in the CoT templates (e.g. COT_TEMPLATE_TL_NORM, COT_TEMPLATE_ANGLE) with per-sample coordinates and measurements.

Parameters:

• template (str) – The template string containing placeholder tokens.

• values_dict (dict) – Mapping of placeholder token to replacement value; each value is stringified before substitution.

Returns:

The template with all placeholders replaced.

Return type:

str