| """Synthetic tool rationales for HanDyVQA, not recorded teacher reasoning. |
| |
| Selection uses only the seed, question ID, and task type. It never reads answers |
| or future observations. Each question has its own RNG, independent of row order. |
| """ |
| import json |
| import random |
|
|
| DEFAULT_SEED = 42 |
| TEMPLATE_VERSION = "handyvqa-tools-v1" |
| FIXED_REASONS = ( |
| "I need to identify the visible entities relevant to the question.", |
| "I need a scene description based on the recognized objects.", |
| "I need temporal observations to assess the relevant motion and state evidence.", |
| ) |
|
|
| |
| RECOG_TEMPLATES = ( |
| "I need to identify the visible entities before I assess the question.", |
| "I will use object recognition to establish which entities are visible.", |
| "I need a list of visible objects to guide further inspection.", |
| "I will identify the objects that could provide evidence for the question.", |
| "I need object recognition before I request more detailed scene evidence.", |
| "I will check which entities appear in the video.", |
| "I need to ground the question in the objects visible in the video.", |
| "I will request object names before I examine their context.", |
| "I need to identify possible subjects of the question in the video.", |
| "I will inspect the video for visible objects before I interpret their roles.", |
| "I need an initial inventory of visible entities.", |
| "I will use recognition to identify objects for further analysis.", |
| "I need to establish which objects the video shows.", |
| "I will request visual object recognition to guide the next observation.", |
| "I need to identify entities without treating the answer options as visual evidence.", |
| "I will check the visible objects before I assess their properties or relationships.", |
| "I need object identities to make the next scene description useful.", |
| "I will identify candidate entities without selecting an answer yet.", |
| "I need to connect the question to entities that the video actually shows.", |
| "I will start with recognition of the visible objects.", |
| ) |
| CAPTION_TEMPLATES = ( |
| "I need a scene description that places the recognized objects in context.", |
| "I will request a description of the scene using the recognized objects.", |
| "I need context beyond the object names before I assess the question.", |
| "I will use captioning to describe the scene around the recognized entities.", |
| "I need a broad description of what the recognized objects are doing or showing.", |
| "I will examine the scene context for the recognized objects.", |
| "I need a description of the objects and their visible relationships.", |
| "I will request scene evidence before I interpret the object list.", |
| "I need to understand how the recognized entities appear in this scene.", |
| "I will use the object list to guide a scene description.", |
| "I need visual context to connect the recognized objects to the question.", |
| "I will ask for a scene description without selecting an answer yet.", |
| "I need more than object identities to assess the visible situation.", |
| "I will request a broad account of the scene and its recognized objects.", |
| "I need a description of the visible context before a closer temporal check.", |
| "I will examine the setting and relationships of the recognized entities.", |
| "I need scene information that object recognition alone does not provide.", |
| "I will request context for the recognized objects without assuming an unseen event.", |
| "I need to connect the object list with the visible situation.", |
| "I will use captioning to establish context for further observations.", |
| ) |
| TRACK_TEMPLATES = ( |
| "I need observations across frames to assess the evidence relevant to the question.", |
| "I will request tracking observations before I select an answer.", |
| "I need a temporal check beyond the broad scene description.", |
| "I will examine the relevant objects across frames.", |
| "I need tracking evidence to assess the visible situation over time.", |
| "I will request observations over time without assuming that a change occurs.", |
| "I need to check what the video supports across its frames.", |
| "I will use tracking to inspect the evidence beyond the scene overview.", |
| "I need observations across frames before I interpret the relevant details.", |
| "I will examine temporal evidence while preserving uncertainty about unseen details.", |
| "I need a closer check of the relevant entities over time.", |
| "I will request tracking evidence to support the answer with visible observations.", |
| "I need to assess which details remain visible across frames.", |
| "I will check the relevant observations over time before drawing a conclusion.", |
| "I need temporal observations rather than assumptions based on a broad description.", |
| "I will use tracking to examine details relevant to the question.", |
| "I need evidence across frames without inferring hidden events.", |
| "I will request a temporal account of the relevant objects.", |
| "I need to inspect the visible evidence over time before choosing an option.", |
| "I will examine the tracked observations without assuming motion or a state change.", |
| ) |
| TOOL_TEMPLATES = (RECOG_TEMPLATES, CAPTION_TEMPLATES, TRACK_TEMPLATES) |
|
|
| |
| TASK_FOCUS = { |
| "state": ( |
| "I will focus on entities whose visible condition could answer the question.", |
| "I need context for the visible condition, without assuming a change.", |
| "I need to distinguish a visible condition from an observed change.", |
| ), |
| "process": ( |
| "I will focus on participants and objects involved in the queried activity.", |
| "I need context for how the queried activity occurs.", |
| "I need the order and manner of actions, where visible.", |
| ), |
| "location": ( |
| "I will focus on the queried entity and possible spatial reference objects.", |
| "I need context for the queried entity's position relative to other objects.", |
| "I need evidence of position and spatial relationships at the relevant time, without assuming movement.", |
| ), |
| "parts": ( |
| "I will focus on the queried object and any visible components.", |
| "I need context for visible parts and their relationship to the whole object.", |
| "I need observations of the queried parts across frames, without assuming that they change.", |
| ), |
| "general": ( |
| "I will focus on entities relevant to the video question.", |
| "I need scene context relevant to the video question.", |
| "I need evidence relevant to the video question, without assuming a change.", |
| ), |
| } |
|
|
|
|
| def template_metadata(template_mode="varied", seed=DEFAULT_SEED): |
| if template_mode not in ("varied", "fixed"): |
| raise ValueError("template_mode must be 'varied' or 'fixed'") |
| if type(seed) is not int: |
| raise ValueError("seed must be an integer") |
| return {"template_mode": template_mode, "template_seed": seed if template_mode == "varied" else None, |
| "template_version": TEMPLATE_VERSION if template_mode == "varied" else "legacy-fixed"} |
|
|
|
|
| def tool_reasons(row, *, template_mode="varied", seed=DEFAULT_SEED): |
| template_metadata(template_mode, seed) |
| if template_mode == "fixed": |
| return FIXED_REASONS |
| task = row.get("question_type") |
| if not task: |
| task = str(row.get("question_id", "")).split("_", 1)[0] |
| if not isinstance(task, str) or task not in TASK_FOCUS: |
| task = "general" |
| rng = random.Random(json.dumps([TEMPLATE_VERSION, seed, row.get("question_id"), task], ensure_ascii=False)) |
| return tuple(rng.choice(candidates) + " " + focus |
| for candidates, focus in zip(TOOL_TEMPLATES, TASK_FOCUS[task])) |
|
|
|
|
| def add_template_arguments(parser): |
| parser.add_argument("--template-mode", choices=("varied", "fixed"), default="varied", |
| help="Select task-aware tool rationales, or retain the three legacy fixed texts") |
| parser.add_argument("--seed", type=int, default=DEFAULT_SEED, |
| help="Seed for per-question template selection (default: 42). Ignored in fixed mode or restore") |
|
|
