Skip to content

compute_diff

autogen.beta.eval.results.diff.compute_diff #

compute_diff(current, baseline, *, strict=True)

Compare current against baseline; see :class:RunDiff and :meth:RunResult.diff.

Source code in autogen/beta/eval/results/diff.py 
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
def compute_diff(current: RunResult, baseline: RunResult, *, strict: bool = True) -> RunDiff:
    """Compare ``current`` against ``baseline``; see :class:`RunDiff` and :meth:`RunResult.diff`."""
    cur = {tr.task.task_id: tr for tr in current.tasks}
    base = {tr.task.task_id: tr for tr in baseline.tasks}

    only_in_current = tuple(sorted(set(cur) - set(base)))
    only_in_baseline = tuple(sorted(set(base) - set(cur)))
    shared = set(cur) & set(base)
    content_changed = tuple(
        sorted(
            tid
            for tid in shared
            if cur[tid].task.inputs != base[tid].task.inputs
            or cur[tid].task.reference_outputs != base[tid].task.reference_outputs
        )
    )
    comparable = sorted(shared - set(content_changed))

    cur_keys = _keys(current.tasks)
    base_keys = _keys(baseline.tasks)
    scorers_only_in_current = tuple(sorted(cur_keys - base_keys))
    scorers_only_in_baseline = tuple(sorted(base_keys - cur_keys))
    shared_keys = sorted(cur_keys & base_keys)

    if strict:
        problems = _problems(
            only_in_current, only_in_baseline, content_changed, scorers_only_in_current, scorers_only_in_baseline
        )
        if problems:
            raise RunsNotComparableError(_message(problems))

    cur_fb = {tid: _by_key(cur[tid]) for tid in comparable}
    base_fb = {tid: _by_key(base[tid]) for tid in comparable}

    pass_rate_deltas: dict[str, tuple[float, float]] = {}
    mean_deltas: dict[str, tuple[float, float]] = {}
    flipped_to_fail: list[tuple[str, str]] = []
    flipped_to_pass: list[tuple[str, str]] = []

    for key in shared_keys:
        base_bools = [s for t in comparable if (s := _bool(base_fb[t].get(key))) is not None]
        cur_bools = [s for t in comparable if (s := _bool(cur_fb[t].get(key))) is not None]
        if base_bools or cur_bools:
            pass_rate_deltas[key] = (_rate(base_bools), _rate(cur_bools))

        base_nums = [float(base_fb[t][key].score) for t in comparable if _is_num(base_fb[t].get(key))]
        cur_nums = [float(cur_fb[t][key].score) for t in comparable if _is_num(cur_fb[t].get(key))]
        if base_nums or cur_nums:
            mean_deltas[key] = (_mean(base_nums), _mean(cur_nums))

        for t in comparable:
            b, c = base_fb[t].get(key), cur_fb[t].get(key)
            if b is None or c is None or not isinstance(b.score, bool) or not isinstance(c.score, bool):
                continue
            if b.score and not c.score:
                flipped_to_fail.append((key, t))
            elif not b.score and c.score:
                flipped_to_pass.append((key, t))

    return RunDiff(
        current_run_id=current.run_id,
        baseline_run_id=baseline.run_id,
        comparable_tasks=tuple(comparable),
        pass_rate_deltas=pass_rate_deltas,
        mean_deltas=mean_deltas,
        flipped_to_fail=tuple(flipped_to_fail),
        flipped_to_pass=tuple(flipped_to_pass),
        only_in_current=only_in_current,
        only_in_baseline=only_in_baseline,
        content_changed=content_changed,
        scorers_only_in_current=scorers_only_in_current,
        scorers_only_in_baseline=scorers_only_in_baseline,
    )