Datasets:

jimchen2
/

clevrer-lite-example

"""CLEVRER-style question generation.
Every question is generated from the *extracted events* of the recorded
trajectory, so questions and answers are grounded in what actually happened.
Question types mirror CLEVRER:
  descriptive    - single/multi-step perception on the first frame
                   ("How many red objects?", "What color is the sphere?")
  explanatory    - causal/temporal reasoning about collision & motion events
                   ("What happened right after ... collided?")
  predictive     - "Will A collide with B by the end of the video?"
  counterfactual - answered by *re-simulating* the scene with one object
                   removed, exactly like the real CLEVRER pipeline.
Each question carries: text, answer, multiple-choice list + answer index,
type, template id, and a simplified functional program.
"""

import numpy as np

from . import config, simulate

_NOTHING = "nothing happened"

_ATTR_ORDER = ["size", "material", "color", "shape"]


# --------------------------------------------------------------------------
# helpers
# --------------------------------------------------------------------------

def _ref(obj, objs):
    """Shortest unambiguous CLEVRER-style reference, e.g. 'the red sphere'."""
    for attrs in (("color", "shape"), ("color",), ("shape",),
                  ("size", "color", "shape"), ("material", "color", "shape"),
                  ("size", "material", "color", "shape")):
        vals = tuple(obj[a] for a in attrs)
        if sum(1 for o in objs if tuple(o[a] for a in attrs) == vals) == 1:
            ordered = [a for a in _ATTR_ORDER if a in attrs]
            return "the " + " ".join(obj[a] for a in ordered)
    return f"the object {obj['idx']}"


def _shuffle_choices(ans, distractors, rng):
    """Build a shuffled choice list; returns (choices, answer_index)."""
    choices = [ans] + [d for d in distractors if d != ans]
    choices = list(dict.fromkeys(choices))          # dedupe, keep order
    perm = rng.permutation(len(choices))
    choices = [choices[i] for i in perm]
    return choices, int(choices.index(ans))


def _numeric_choices(ans, lo, hi, rng, k=6):
    vals = {ans}
    pool = [v for v in range(lo, hi + 1) if v != ans]
    if pool:
        vals |= set(int(x) for x in rng.choice(pool, size=min(k - 1, len(pool)),
                                               replace=False))
    return [str(v) for v in sorted(vals)]


def _q(question, answer, choices, answer_index, qtype, template, program):
    answer = str(answer)
    if answer_index is None:
        answer_index = choices.index(answer)
    return {"question": question, "answer": answer, "choices": choices,
            "answer_index": answer_index, "question_type": qtype,
            "template": template, "program": program}


def _col_choices(ans, rng):
    names = list(config.COLOR_NAMES)
    perm = rng.permutation(len(names))
    choices = [names[i] for i in perm]
    return choices, choices.index(ans)


# --------------------------------------------------------------------------
# descriptive questions (about the first frame)
# --------------------------------------------------------------------------

def _descriptive(objs, ev, rng):
    out = []
    n = len(objs)
    pos0 = {o["idx"]: o["position"] for o in objs}

    out.append(_q("How many objects are there?", str(n),
                  _numeric_choices(n, 0, max(6, n), rng), None,
                  "descriptive", "count_all",
                  [{"op": "count"}]))

    # count by attribute
    for key, noun in (("color", "objects"), ("shape", "objects"),
                      ("material", "objects")):
        values = sorted({o[key] for o in objs})
        for v in values:
            cnt = sum(1 for o in objs if o[key] == v)
            phrase = f"How many {v} {noun} are there?"
            prog = [{"op": f"filter_{key}", "value": v}, {"op": "count"}]
            ch = _numeric_choices(cnt, 0, max(6, n), rng)
            out.append(_q(phrase, str(cnt), ch, None, "descriptive",
                          f"count_{key}", prog))

    # query color of a unique shape
    for sh in config.SHAPES:
        same = [o for o in objs if o["shape"] == sh]
        if len(same) == 1:
            o = same[0]
            ch, ai = _col_choices(o["color"], rng)
            out.append(_q(f"What color is the {sh}?", o["color"], ch, ai,
                          "descriptive", "query_color",
                          [{"op": "filter_shape", "value": sh},
                           {"op": "query_color"}]))

    # query shape of a unique color
    for col in config.COLOR_NAMES:
        same = [o for o in objs if o["color"] == col]
        if len(same) == 1:
            o = same[0]
            ch = list(config.SHAPES)
            perm = rng.permutation(len(ch))
            ch = [ch[i] for i in perm]
            out.append(_q(f"What shape is the {col} object?", o["shape"], ch,
                          list(ch).index(o["shape"]), "descriptive",
                          "query_shape",
                          [{"op": "filter_color", "value": col},
                           {"op": "query_shape"}]))

    # query material of a unique color+shape
    for o in objs:
        same = [x for x in objs if x["color"] == o["color"] and x["shape"] == o["shape"]]
        if len(same) == 1:
            ch = list(config.MATERIALS)
            perm = rng.permutation(len(ch))
            ch = [ch[i] for i in perm]
            out.append(_q(f"What material is the {o['color']} {o['shape']} made of?",
                          o["material"], ch, ch.index(o["material"]),
                          "descriptive", "query_material",
                          [{"op": "filter_color", "value": o["color"]},
                           {"op": "filter_shape", "value": o["shape"]},
                           {"op": "query_material"}]))

    # spatial relations (margin so 'left of' is unambiguous, exactly one match)
    margin = 0.6
    dirs = {
        "left of":   lambda a, b: (b[0] - a[0]) >= margin,   # a is left of b
        "right of":  lambda a, b: (a[0] - b[0]) >= margin,
        "behind":    lambda a, b: (a[1] - b[1]) >= margin,   # +y is far/back
        "in front of": lambda a, b: (b[1] - a[1]) >= margin,
    }
    for rb in objs:
        for name, fn in dirs.items():
            cands = [o for o in objs if o["idx"] != rb["idx"] and
                     fn(pos0[o["idx"]], pos0[rb["idx"]])]
            if len(cands) == 1:
                o = cands[0]
                ch, ai = _col_choices(o["color"], rng)
                out.append(_q(f"What color is the object {name} {_ref(rb, objs)}?",
                              o["color"], ch, ai, "descriptive", "query_color_rel",
                              [{"op": "filter_color", "value": rb["color"]},
                               {"op": "relate", "direction": name.replace(" ", "_")},
                               {"op": "query_color"}]))
    return out


# --------------------------------------------------------------------------
# explanatory questions (causal / temporal)
# --------------------------------------------------------------------------

def _coll_desc(objs, a, b):
    return f"{_ref(a, objs)} collided with {_ref(b, objs)}"


def _explanatory(objs, ev, rng):
    out = []
    cev = ev["collisions"]
    idx_of = {o["idx"]: o for o in objs}
    refc = {o["idx"]: _ref(o, objs) for o in objs}

    if not cev:
        return out

    def ev_desc(e):
        a, b = e["objects"]
        return _coll_desc(objs, idx_of[a], idx_of[b])

    # what happened right after / before a collision
    for k, e in enumerate(cev):
        for direction, other in (("after", k + 1), ("before", k - 1)):
            ans = ev_desc(cev[other]) if 0 <= other < len(cev) else _NOTHING
            distractors = [ev_desc(x) for i, x in enumerate(cev)
                           if i not in (k, other)]
            for (t, i, ax, s) in ev["wall_hits"][:40]:
                d = f"{refc[i]} collided with the wall"
                if d not in distractors:
                    distractors.append(d)
            for o, sts in ev["starts"].items():
                if o in refc and sts:
                    d = f"{refc[o]} started moving"
                    if d not in distractors:
                        distractors.append(d)
            if ans != _NOTHING and _NOTHING not in distractors:
                distractors.append(_NOTHING)
            if len(distractors) < 3:
                continue
            rng.shuffle(distractors)
            distractors = distractors[:3]
            a, b = e["objects"]
            ch, ai = _shuffle_choices(ans, distractors, rng)
            out.append(_q(f"What happened right {direction} "
                          f"{ev_desc(e)}?", ans, ch, ai, "explanatory",
                          f"what_happened_{direction}",
                          [{"op": "find_collision", "objects": [a, b]},
                           {"op": f"query_{direction}_event"}]))

    # what caused an object to start moving
    for o_idx, sts in ev["starts"].items():
        if o_idx not in idx_of:
            continue
        for s in sts:
            cause = None
            for e in cev:
                a, b = e["objects"]
                if o_idx in (a, b) and abs(e["frame"] - s) <= config.CAUSE_WINDOW_FRAMES:
                    cause = b if a == o_idx else a
                    break
            if cause is None:
                continue
            ans = f"the collision with {refc[cause]}"
            others = [p for e in cev for p in e["objects"]
                      if p not in (o_idx, cause) and p in refc]
            distractors = [f"the collision with {refc[p]}" for p in dict.fromkeys(others)]
            distractors += ["it was pushed by the wall", "nothing - it was already moving"]
            rng.shuffle(distractors)
            ch, ai = _shuffle_choices(ans, distractors[:3], rng)
            out.append(_q(f"What caused {refc[o_idx]} to start moving?", ans, ch, ai,
                          "explanatory", "cause_of_motion",
                          [{"op": "filter_object", "idx": o_idx},
                           {"op": "query_motion_cause"}]))

    # which object started moving right before / after a collision
    for e in cev:
        a, b = e["objects"]
        t = e["frame"]
        for direction, lo, hi, pick in (("after", t + 1, t + 12, min),
                                        ("before", max(1, t - 12), t, max)):
            cands = [(s, o) for o, ss in ev["starts"].items()
                     if o not in (a, b) and o in idx_of for s in ss
                     if lo <= s <= hi]
            if not cands:
                continue
            _, ans_o = pick(cands)
            ans = refc[ans_o]
            others = [o for o in objs if o["idx"] not in (ans_o, a, b)]
            distractors = [_ref(o, objs) for o in others]
            distractors.append("none of them")
            rng.shuffle(distractors)
            ch, ai = _shuffle_choices(ans, distractors[:3], rng)
            out.append(_q(f"Which object started moving right {direction} "
                          f"{ev_desc(e)}?", ans, ch, ai, "explanatory",
                          f"who_started_{direction}",
                          [{"op": "find_collision", "objects": [a, b]},
                           {"op": f"query_{direction}_motion_start"}]))

    # how many objects were moving when a collision happened
    for e in cev:
        a, b = e["objects"]
        t = min(e["frame"], ev["moving"].shape[0] - 1)
        cnt = int(ev["moving"][t].sum())
        ch = _numeric_choices(cnt, 0, len(objs), rng)
        out.append(_q(f"How many objects were moving when {ev_desc(e)}?",
                      str(cnt), ch, None, "explanatory", "count_moving_at",
                      [{"op": "find_collision", "objects": [a, b]},
                       {"op": "count_moving_at_collision"}]))

    # total number of collisions
    if len(cev) >= 2:
        ch = _numeric_choices(len(cev), 0, max(6, len(cev) + 1), rng)
        out.append(_q("How many collisions happened in the video?",
                      str(len(cev)), ch, None, "explanatory",
                      "count_collisions",
                      [{"op": "filter_event", "type": "collision"}, {"op": "count"}]))
    return out


# --------------------------------------------------------------------------
# predictive questions (binary, grounded in the full recorded trajectory)
# --------------------------------------------------------------------------

def _predictive(objs, ev, rng):
    out = []
    cev = ev["collisions"]
    idx_of = {o["idx"]: o for o in objs}
    pairs_collided = {tuple(sorted(e["objects"])) for e in cev}
    all_pairs = [tuple(sorted((a["idx"], b["idx"])))
                 for i, a in enumerate(objs) for b in objs[i + 1:]]

    def binary_q(pair, ans_yes, template):
        a, b = pair
        ch = ["yes", "no"]
        rng.shuffle(ch)
        ans = "yes" if ans_yes else "no"
        return _q(f"Will {_ref(idx_of[a], objs)} collide with "
                  f"{_ref(idx_of[b], objs)} by the end of the video?",
                  ans, ch, ch.index(ans), "predictive", template,
                  [{"op": "find_collision", "objects": [a, b]},
                   {"op": "query_existence"}])

    yes_pairs = [p for p in pairs_collided]
    no_pairs = [p for p in all_pairs if p not in pairs_collided]
    if yes_pairs:
        p = yes_pairs[int(rng.integers(len(yes_pairs)))]
        out.append(binary_q(p, True, "will_collide"))
    if no_pairs:
        p = no_pairs[int(rng.integers(len(no_pairs)))]
        out.append(binary_q(p, False, "will_collide"))

    # after colliding with b, will a collide with something else?
    for e in cev:
        a, b = e["objects"]
        later = [x for x in cev if x is not e and a in x["objects"]
                 and x["frame"] > e["frame"]]
        if not later and len(cev) < 2:
            continue
        ans = "yes" if later else "no"
        ch = ["yes", "no"]
        rng.shuffle(ch)
        out.append(_q(f"After {_coll_desc(objs, idx_of[a], idx_of[b])}, "
                      f"will {_ref(idx_of[a], objs)} collide with something else?",
                      ans, ch, ch.index(ans), "predictive", "will_collide_again",
                      [{"op": "find_collision", "objects": [a, b]},
                       {"op": "query_subsequent_collision"}]))
    return out


# --------------------------------------------------------------------------
# counterfactual questions (re-simulate the scene with one object removed)
# --------------------------------------------------------------------------

def _counterfactual(objs, ev, rng, num_frames):
    out = []
    cev = ev["collisions"]
    if not cev or len(objs) < 3:
        return out
    idx_of = {o["idx"]: o for o in objs}
    refc = {o["idx"]: _ref(o, objs) for o in objs}

    cache = {}

    def cf_events(z):
        if z not in cache:
            cache[z] = simulate.resimulate_without(
                [dict(s) for s in objs], z, num_frames)
        return cache[z]

    def binary(ans_yes, question, template, program):
        ch = ["yes", "no"]
        rng.shuffle(ch)
        ans = "yes" if ans_yes else "no"
        out.append(_q(question, ans, ch, ch.index(ans), "counterfactual",
                      template, program))

    # would a still collide with b without z?
    for e in cev:
        a, b = e["objects"]
        others = [o["idx"] for o in objs if o["idx"] not in (a, b)]
        for z in others:
            cf = cf_events(z)
            still = any(set(x["objects"]) == {a, b} for x in cf["collisions"])
            binary(still,
                   f"If {refc[z]} were removed, would {refc[a]} still collide "
                   f"with {refc[b]}?",
                   "cf_collision",
                   [{"op": "remove_object", "idx": z},
                    {"op": "find_collision", "objects": [a, b]},
                    {"op": "query_existence"}])

    # would o still start moving without z (z can be its collision partner)?
    for o_idx, sts in ev["starts"].items():
        if not sts or o_idx not in idx_of:
            continue
        partners = set()
        for e in cev:
            if o_idx in e["objects"]:
                partners |= {p for p in e["objects"] if p != o_idx}
        others = list(partners) + [o["idx"] for o in objs
                                   if o["idx"] not in {o_idx} | partners]
        for z in dict.fromkeys(others):
            cf = cf_events(z)
            still = (o_idx in cf["initially_moving"]) or bool(cf["starts"].get(o_idx))
            binary(still,
                   f"If {refc[z]} were removed, would {refc[o_idx]} still "
                   f"start moving?",
                   "cf_motion",
                   [{"op": "remove_object", "idx": z},
                    {"op": "filter_object", "idx": o_idx},
                    {"op": "query_motion_start"}])
    return out


# --------------------------------------------------------------------------
# orchestration
# --------------------------------------------------------------------------

def build_questions(spec, ev, rng, num_frames, max_questions=8):
    """Return a list of question dicts sampled across the 4 CLEVRER types."""
    objs = [s for s in spec if not s.get("is_probe")]
    probe_idx = {s["idx"] for s in spec if s.get("is_probe")}

    if probe_idx:  # drop probe objects from question-relevant events
        ev = {
            "collisions": [e for e in ev["collisions"] if not probe_idx & set(e["objects"])],
            "starts": {k: v for k, v in ev["starts"].items() if k not in probe_idx},
            "stops": {k: v for k, v in ev["stops"].items() if k not in probe_idx},
            "initially_moving": [k for k in ev["initially_moving"] if k not in probe_idx],
            "moving": ev["moving"],
            "speeds": ev["speeds"],
            "wall_hits": [w for w in ev["wall_hits"] if w[1] not in probe_idx],
        }

    pools = {
        "descriptive": _descriptive(objs, ev, rng),
        "explanatory": _explanatory(objs, ev, rng),
        "predictive": _predictive(objs, ev, rng),
        "counterfactual": _counterfactual(objs, ev, rng, num_frames),
    }

    quotas = [("descriptive", 3), ("explanatory", 2),
              ("predictive", 1), ("counterfactual", 2)]
    picked, seen = [], set()

    def take(pool, k):
        pool = [q for q in pool if q["question"] not in seen]
        rng.shuffle(pool)
        for q in pool[:k]:
            seen.add(q["question"])
            picked.append(q)

    for qtype, k in quotas:
        take(pools[qtype], k)
    leftover = [q for pool in pools.values() for q in pool
                if q["question"] not in seen]
    rng.shuffle(leftover)
    for q in leftover:
        if len(picked) >= max_questions:
            break
        seen.add(q["question"])
        picked.append(q)

    rng.shuffle(picked)
    return picked

	"""CLEVRER-style question generation.

	Every question is generated from the extracted events of the recorded
	trajectory, so questions and answers are grounded in what actually happened.
	Question types mirror CLEVRER:

	descriptive - single/multi-step perception on the first frame
	("How many red objects?", "What color is the sphere?")
	explanatory - causal/temporal reasoning about collision & motion events
	("What happened right after ... collided?")
	predictive - "Will A collide with B by the end of the video?"
	counterfactual - answered by re-simulating the scene with one object
	removed, exactly like the real CLEVRER pipeline.

	Each question carries: text, answer, multiple-choice list + answer index,
	type, template id, and a simplified functional program.
	"""

	import numpy as np

	from . import config, simulate

	_NOTHING = "nothing happened"

	_ATTR_ORDER = ["size", "material", "color", "shape"]


	# --------------------------------------------------------------------------
	# helpers
	# --------------------------------------------------------------------------

	def _ref(obj, objs):
	"""Shortest unambiguous CLEVRER-style reference, e.g. 'the red sphere'."""
	for attrs in (("color", "shape"), ("color",), ("shape",),
	("size", "color", "shape"), ("material", "color", "shape"),
	("size", "material", "color", "shape")):
	vals = tuple(obj[a] for a in attrs)
	if sum(1 for o in objs if tuple(o[a] for a in attrs) == vals) == 1:
	ordered = [a for a in _ATTR_ORDER if a in attrs]
	return "the " + " ".join(obj[a] for a in ordered)
	return f"the object {obj['idx']}"


	def _shuffle_choices(ans, distractors, rng):
	"""Build a shuffled choice list; returns (choices, answer_index)."""
	choices = [ans] + [d for d in distractors if d != ans]
	choices = list(dict.fromkeys(choices)) # dedupe, keep order
	perm = rng.permutation(len(choices))
	choices = [choices[i] for i in perm]
	return choices, int(choices.index(ans))


	def _numeric_choices(ans, lo, hi, rng, k=6):
	vals = {ans}
	pool = [v for v in range(lo, hi + 1) if v != ans]
	if pool:
	vals \|= set(int(x) for x in rng.choice(pool, size=min(k - 1, len(pool)),
	replace=False))
	return [str(v) for v in sorted(vals)]


	def _q(question, answer, choices, answer_index, qtype, template, program):
	answer = str(answer)
	if answer_index is None:
	answer_index = choices.index(answer)
	return {"question": question, "answer": answer, "choices": choices,
	"answer_index": answer_index, "question_type": qtype,
	"template": template, "program": program}


	def _col_choices(ans, rng):
	names = list(config.COLOR_NAMES)
	perm = rng.permutation(len(names))
	choices = [names[i] for i in perm]
	return choices, choices.index(ans)


	# --------------------------------------------------------------------------
	# descriptive questions (about the first frame)
	# --------------------------------------------------------------------------

	def _descriptive(objs, ev, rng):
	out = []
	n = len(objs)
	pos0 = {o["idx"]: o["position"] for o in objs}

	out.append(_q("How many objects are there?", str(n),
	_numeric_choices(n, 0, max(6, n), rng), None,
	"descriptive", "count_all",
	[{"op": "count"}]))

	# count by attribute
	for key, noun in (("color", "objects"), ("shape", "objects"),
	("material", "objects")):
	values = sorted({o[key] for o in objs})
	for v in values:
	cnt = sum(1 for o in objs if o[key] == v)
	phrase = f"How many {v} {noun} are there?"
	prog = [{"op": f"filter_{key}", "value": v}, {"op": "count"}]
	ch = _numeric_choices(cnt, 0, max(6, n), rng)
	out.append(_q(phrase, str(cnt), ch, None, "descriptive",
	f"count_{key}", prog))

	# query color of a unique shape
	for sh in config.SHAPES:
	same = [o for o in objs if o["shape"] == sh]
	if len(same) == 1:
	o = same[0]
	ch, ai = _col_choices(o["color"], rng)
	out.append(_q(f"What color is the {sh}?", o["color"], ch, ai,
	"descriptive", "query_color",
	[{"op": "filter_shape", "value": sh},
	{"op": "query_color"}]))

	# query shape of a unique color
	for col in config.COLOR_NAMES:
	same = [o for o in objs if o["color"] == col]
	if len(same) == 1:
	o = same[0]
	ch = list(config.SHAPES)
	perm = rng.permutation(len(ch))
	ch = [ch[i] for i in perm]
	out.append(_q(f"What shape is the {col} object?", o["shape"], ch,
	list(ch).index(o["shape"]), "descriptive",
	"query_shape",
	[{"op": "filter_color", "value": col},
	{"op": "query_shape"}]))

	# query material of a unique color+shape
	for o in objs:
	same = [x for x in objs if x["color"] == o["color"] and x["shape"] == o["shape"]]
	if len(same) == 1:
	ch = list(config.MATERIALS)
	perm = rng.permutation(len(ch))
	ch = [ch[i] for i in perm]
	out.append(_q(f"What material is the {o['color']} {o['shape']} made of?",
	o["material"], ch, ch.index(o["material"]),
	"descriptive", "query_material",
	[{"op": "filter_color", "value": o["color"]},
	{"op": "filter_shape", "value": o["shape"]},
	{"op": "query_material"}]))

	# spatial relations (margin so 'left of' is unambiguous, exactly one match)
	margin = 0.6
	dirs = {
	"left of": lambda a, b: (b[0] - a[0]) >= margin, # a is left of b
	"right of": lambda a, b: (a[0] - b[0]) >= margin,
	"behind": lambda a, b: (a[1] - b[1]) >= margin, # +y is far/back
	"in front of": lambda a, b: (b[1] - a[1]) >= margin,
	}
	for rb in objs:
	for name, fn in dirs.items():
	cands = [o for o in objs if o["idx"] != rb["idx"] and
	fn(pos0[o["idx"]], pos0[rb["idx"]])]
	if len(cands) == 1:
	o = cands[0]
	ch, ai = _col_choices(o["color"], rng)
	out.append(_q(f"What color is the object {name} {_ref(rb, objs)}?",
	o["color"], ch, ai, "descriptive", "query_color_rel",
	[{"op": "filter_color", "value": rb["color"]},
	{"op": "relate", "direction": name.replace(" ", "_")},
	{"op": "query_color"}]))
	return out


	# --------------------------------------------------------------------------
	# explanatory questions (causal / temporal)
	# --------------------------------------------------------------------------

	def _coll_desc(objs, a, b):
	return f"{_ref(a, objs)} collided with {_ref(b, objs)}"


	def _explanatory(objs, ev, rng):
	out = []
	cev = ev["collisions"]
	idx_of = {o["idx"]: o for o in objs}
	refc = {o["idx"]: _ref(o, objs) for o in objs}

	if not cev:
	return out

	def ev_desc(e):
	a, b = e["objects"]
	return _coll_desc(objs, idx_of[a], idx_of[b])

	# what happened right after / before a collision
	for k, e in enumerate(cev):
	for direction, other in (("after", k + 1), ("before", k - 1)):
	ans = ev_desc(cev[other]) if 0 <= other < len(cev) else _NOTHING
	distractors = [ev_desc(x) for i, x in enumerate(cev)
	if i not in (k, other)]
	for (t, i, ax, s) in ev["wall_hits"][:40]:
	d = f"{refc[i]} collided with the wall"
	if d not in distractors:
	distractors.append(d)
	for o, sts in ev["starts"].items():
	if o in refc and sts:
	d = f"{refc[o]} started moving"
	if d not in distractors:
	distractors.append(d)
	if ans != _NOTHING and _NOTHING not in distractors:
	distractors.append(_NOTHING)
	if len(distractors) < 3:
	continue
	rng.shuffle(distractors)
	distractors = distractors[:3]
	a, b = e["objects"]
	ch, ai = _shuffle_choices(ans, distractors, rng)
	out.append(_q(f"What happened right {direction} "
	f"{ev_desc(e)}?", ans, ch, ai, "explanatory",
	f"what_happened_{direction}",
	[{"op": "find_collision", "objects": [a, b]},
	{"op": f"query_{direction}_event"}]))

	# what caused an object to start moving
	for o_idx, sts in ev["starts"].items():
	if o_idx not in idx_of:
	continue
	for s in sts:
	cause = None
	for e in cev:
	a, b = e["objects"]
	if o_idx in (a, b) and abs(e["frame"] - s) <= config.CAUSE_WINDOW_FRAMES:
	cause = b if a == o_idx else a
	break
	if cause is None:
	continue
	ans = f"the collision with {refc[cause]}"
	others = [p for e in cev for p in e["objects"]
	if p not in (o_idx, cause) and p in refc]
	distractors = [f"the collision with {refc[p]}" for p in dict.fromkeys(others)]
	distractors += ["it was pushed by the wall", "nothing - it was already moving"]
	rng.shuffle(distractors)
	ch, ai = _shuffle_choices(ans, distractors[:3], rng)
	out.append(_q(f"What caused {refc[o_idx]} to start moving?", ans, ch, ai,
	"explanatory", "cause_of_motion",
	[{"op": "filter_object", "idx": o_idx},
	{"op": "query_motion_cause"}]))

	# which object started moving right before / after a collision
	for e in cev:
	a, b = e["objects"]
	t = e["frame"]
	for direction, lo, hi, pick in (("after", t + 1, t + 12, min),
	("before", max(1, t - 12), t, max)):
	cands = [(s, o) for o, ss in ev["starts"].items()
	if o not in (a, b) and o in idx_of for s in ss
	if lo <= s <= hi]
	if not cands:
	continue
	_, ans_o = pick(cands)
	ans = refc[ans_o]
	others = [o for o in objs if o["idx"] not in (ans_o, a, b)]
	distractors = [_ref(o, objs) for o in others]
	distractors.append("none of them")
	rng.shuffle(distractors)
	ch, ai = _shuffle_choices(ans, distractors[:3], rng)
	out.append(_q(f"Which object started moving right {direction} "
	f"{ev_desc(e)}?", ans, ch, ai, "explanatory",
	f"who_started_{direction}",
	[{"op": "find_collision", "objects": [a, b]},
	{"op": f"query_{direction}_motion_start"}]))

	# how many objects were moving when a collision happened
	for e in cev:
	a, b = e["objects"]
	t = min(e["frame"], ev["moving"].shape[0] - 1)
	cnt = int(ev["moving"][t].sum())
	ch = _numeric_choices(cnt, 0, len(objs), rng)
	out.append(_q(f"How many objects were moving when {ev_desc(e)}?",
	str(cnt), ch, None, "explanatory", "count_moving_at",
	[{"op": "find_collision", "objects": [a, b]},
	{"op": "count_moving_at_collision"}]))

	# total number of collisions
	if len(cev) >= 2:
	ch = _numeric_choices(len(cev), 0, max(6, len(cev) + 1), rng)
	out.append(_q("How many collisions happened in the video?",
	str(len(cev)), ch, None, "explanatory",
	"count_collisions",
	[{"op": "filter_event", "type": "collision"}, {"op": "count"}]))
	return out


	# --------------------------------------------------------------------------
	# predictive questions (binary, grounded in the full recorded trajectory)
	# --------------------------------------------------------------------------

	def _predictive(objs, ev, rng):
	out = []
	cev = ev["collisions"]
	idx_of = {o["idx"]: o for o in objs}
	pairs_collided = {tuple(sorted(e["objects"])) for e in cev}
	all_pairs = [tuple(sorted((a["idx"], b["idx"])))
	for i, a in enumerate(objs) for b in objs[i + 1:]]

	def binary_q(pair, ans_yes, template):
	a, b = pair
	ch = ["yes", "no"]
	rng.shuffle(ch)
	ans = "yes" if ans_yes else "no"
	return _q(f"Will {_ref(idx_of[a], objs)} collide with "
	f"{_ref(idx_of[b], objs)} by the end of the video?",
	ans, ch, ch.index(ans), "predictive", template,
	[{"op": "find_collision", "objects": [a, b]},
	{"op": "query_existence"}])

	yes_pairs = [p for p in pairs_collided]
	no_pairs = [p for p in all_pairs if p not in pairs_collided]
	if yes_pairs:
	p = yes_pairs[int(rng.integers(len(yes_pairs)))]
	out.append(binary_q(p, True, "will_collide"))
	if no_pairs:
	p = no_pairs[int(rng.integers(len(no_pairs)))]
	out.append(binary_q(p, False, "will_collide"))

	# after colliding with b, will a collide with something else?
	for e in cev:
	a, b = e["objects"]
	later = [x for x in cev if x is not e and a in x["objects"]
	and x["frame"] > e["frame"]]
	if not later and len(cev) < 2:
	continue
	ans = "yes" if later else "no"
	ch = ["yes", "no"]
	rng.shuffle(ch)
	out.append(_q(f"After {_coll_desc(objs, idx_of[a], idx_of[b])}, "
	f"will {_ref(idx_of[a], objs)} collide with something else?",
	ans, ch, ch.index(ans), "predictive", "will_collide_again",
	[{"op": "find_collision", "objects": [a, b]},
	{"op": "query_subsequent_collision"}]))
	return out


	# --------------------------------------------------------------------------
	# counterfactual questions (re-simulate the scene with one object removed)
	# --------------------------------------------------------------------------

	def _counterfactual(objs, ev, rng, num_frames):
	out = []
	cev = ev["collisions"]
	if not cev or len(objs) < 3:
	return out
	idx_of = {o["idx"]: o for o in objs}
	refc = {o["idx"]: _ref(o, objs) for o in objs}

	cache = {}

	def cf_events(z):
	if z not in cache:
	cache[z] = simulate.resimulate_without(
	[dict(s) for s in objs], z, num_frames)
	return cache[z]

	def binary(ans_yes, question, template, program):
	ch = ["yes", "no"]
	rng.shuffle(ch)
	ans = "yes" if ans_yes else "no"
	out.append(_q(question, ans, ch, ch.index(ans), "counterfactual",
	template, program))

	# would a still collide with b without z?
	for e in cev:
	a, b = e["objects"]
	others = [o["idx"] for o in objs if o["idx"] not in (a, b)]
	for z in others:
	cf = cf_events(z)
	still = any(set(x["objects"]) == {a, b} for x in cf["collisions"])
	binary(still,
	f"If {refc[z]} were removed, would {refc[a]} still collide "
	f"with {refc[b]}?",
	"cf_collision",
	[{"op": "remove_object", "idx": z},
	{"op": "find_collision", "objects": [a, b]},
	{"op": "query_existence"}])

	# would o still start moving without z (z can be its collision partner)?
	for o_idx, sts in ev["starts"].items():
	if not sts or o_idx not in idx_of:
	continue
	partners = set()
	for e in cev:
	if o_idx in e["objects"]:
	partners \|= {p for p in e["objects"] if p != o_idx}
	others = list(partners) + [o["idx"] for o in objs
	if o["idx"] not in {o_idx} \| partners]
	for z in dict.fromkeys(others):
	cf = cf_events(z)
	still = (o_idx in cf["initially_moving"]) or bool(cf["starts"].get(o_idx))
	binary(still,
	f"If {refc[z]} were removed, would {refc[o_idx]} still "
	f"start moving?",
	"cf_motion",
	[{"op": "remove_object", "idx": z},
	{"op": "filter_object", "idx": o_idx},
	{"op": "query_motion_start"}])
	return out


	# --------------------------------------------------------------------------
	# orchestration
	# --------------------------------------------------------------------------

	def build_questions(spec, ev, rng, num_frames, max_questions=8):
	"""Return a list of question dicts sampled across the 4 CLEVRER types."""
	objs = [s for s in spec if not s.get("is_probe")]
	probe_idx = {s["idx"] for s in spec if s.get("is_probe")}

	if probe_idx: # drop probe objects from question-relevant events
	ev = {
	"collisions": [e for e in ev["collisions"] if not probe_idx & set(e["objects"])],
	"starts": {k: v for k, v in ev["starts"].items() if k not in probe_idx},
	"stops": {k: v for k, v in ev["stops"].items() if k not in probe_idx},
	"initially_moving": [k for k in ev["initially_moving"] if k not in probe_idx],
	"moving": ev["moving"],
	"speeds": ev["speeds"],
	"wall_hits": [w for w in ev["wall_hits"] if w[1] not in probe_idx],
	}

	pools = {
	"descriptive": _descriptive(objs, ev, rng),
	"explanatory": _explanatory(objs, ev, rng),
	"predictive": _predictive(objs, ev, rng),
	"counterfactual": _counterfactual(objs, ev, rng, num_frames),
	}

	quotas = [("descriptive", 3), ("explanatory", 2),
	("predictive", 1), ("counterfactual", 2)]
	picked, seen = [], set()

	def take(pool, k):
	pool = [q for q in pool if q["question"] not in seen]
	rng.shuffle(pool)
	for q in pool[:k]:
	seen.add(q["question"])
	picked.append(q)

	for qtype, k in quotas:
	take(pools[qtype], k)
	leftover = [q for pool in pools.values() for q in pool
	if q["question"] not in seen]
	rng.shuffle(leftover)
	for q in leftover:
	if len(picked) >= max_questions:
	break
	seen.add(q["question"])
	picked.append(q)

	rng.shuffle(picked)
	return picked