| |
| """Precision and recall per role of a role predictor against the labels. |
| |
| python3 scripts/evaluate_roles.py predictions.jsonl [--projects kiwi,mall-swarm] |
| |
| predictions.jsonl holds one {"id": "<project>/<qualified name>", "role": "<role>"} per line; |
| a missing id counts as "none". "none" predicted is no prediction: a miss counts against |
| the labelled role, not as a false positive of "none". See results/agentkit-roles.predictions.jsonl. |
| """ |
| import argparse |
|
|
| from common import ROLES, data, read_jsonl |
|
|
| p = argparse.ArgumentParser() |
| p.add_argument("predictions") |
| p.add_argument("--projects", default="") |
| a = p.parse_args() |
| only = {x for x in a.projects.split(",") if x} |
| labels = [l for l in read_jsonl(data("roles", "labels.jsonl")) if not only or l["project"] in only] |
| pred = {r["id"]: r["role"] for r in read_jsonl(a.predictions)} |
| c = {r: {"tp": 0, "fp": 0, "fn": 0, "n": 0} for r in ROLES} |
| for l in labels: |
| got = pred.get(l["id"], "none") |
| c[l["role"]]["n"] += 1 |
| if got == l["role"]: |
| c[l["role"]]["tp"] += 1 |
| else: |
| c[l["role"]]["fn"] += 1 |
| if got != "none": |
| c[got]["fp"] += 1 |
| pct = lambda n, d: f"{100 * n / d:5.1f} %" if d else " — " |
| print(f"{'role':<12} {'precision':>10} {'recall':>10} {'labelled':>9}") |
| for r in ROLES: |
| x = c[r] |
| if x["n"] or x["fp"]: |
| prec = " — " if r == "none" else pct(x["tp"], x["tp"] + x["fp"]) |
| print(f"{r:<12} {prec:>10} {pct(x['tp'], x['tp'] + x['fn']):>10} {x['n']:>9}") |
| right = sum(x["tp"] for x in c.values()) |
| print(f"accuracy {pct(right, len(labels)).strip()} ({right}/{len(labels)})") |
|
|
