bench: save per-decision probabilities and add them to the eval-results PR

score_with_jobs.py already scores every decision but keeps only the summary. This PR:

- writes `predictions.jsonl` next to `results.json`: one line per decision with case id, workflow, question id, type and the normalized probabilities over the gold options (the vector `score_one` scores, after the yes/no to true/false mapping);
- has `submit_results.py` add that file to the model-repo PR as `eval/typed-decisions/predictions.jsonl` when it is present next to the results file (`--predictions` points elsewhere).

Scoring is unchanged. Replaying a published Jev run (OpenDecider's whole-case file, which their benchmarks README lists as Jev 1.13 through TypeSafe's API) through both versions with vLLM, HTTP and `datasets` mocked gives a byte-identical `results.json` (accuracy 0.7305), and re-scoring `predictions.jsonl` gives the same accuracy, KL, Brier and ECE.

The tie rule (first-listed option vs split credit) is left for a separate change; with the probabilities saved it can be re-applied to any run.

Context: https://huggingface.co/posts/tardellirs/145368791310880

bench/score_with_jobs.py

@@ -6,7 +6,8 @@ Brier and ECE over the full distributions.
 
 Usage (inside the job, after setup): python bench/score_with_jobs.py MODEL_ID
 Optional env: SERVE_FLAGS (extra `vllm serve` flags).
-Results are printed and saved to results.json.
+Results are printed and saved to results.json; the normalized probabilities of every
+decision go to predictions.jsonl, one line per decision keyed on case id and question id.
 """
 import json
 import math
@@ -38,7 +39,7 @@ def score_one(pred, gold):
     am_g = g.index(max(g))
     kl = sum(gi * math.log(gi / max(pi, 1e-9)) for gi, pi in zip(g, p) if gi > 0)
     brier = sum((pi - gi) ** 2 for pi, gi in zip(p, g))
-    return am_p == am_g, kl, brier, max(p)
+    return am_p == am_g, kl, brier, max(p), dict(zip(opts, p))
 
 
 def main():
@@ -60,7 +61,7 @@ def main():
     print(f"[bench] server ready for {model}", flush=True)
 
     ds = load_dataset("LocalLLaMA/typed-decisions", "all", split="test")
-    accs, kls, briers, confs = [], [], [], []
+    accs, kls, briers, confs, preds = [], [], [], [], []
     for i, raw in enumerate(ds):
         payload = {
             "model": model,
@@ -72,7 +73,9 @@ def main():
         answers = r.json()["answers"]
         gold = json.loads(raw["gold"])
         for qid, gq in gold.items():
-            ok, kl, brier, conf = score_one(answers.get(qid, {"probabilities": {}}), gq)
+            ok, kl, brier, conf, probs = score_one(answers.get(qid, {"probabilities": {}}), gq)
+            preds.append({"id": raw["id"], "workflow": raw["workflow"], "question": qid,
+                          "type": gq["type"], "probabilities": probs})
             accs.append(ok)
             kls.append(kl)
             briers.append(brier)
@@ -105,6 +108,9 @@ def main():
     }
     with open("results.json", "w") as f:
         json.dump(summary, f, indent=2)
+    with open("predictions.jsonl", "w") as f:
+        for p in preds:
+            f.write(json.dumps(p) + "\n")
     print(json.dumps(summary, indent=2), flush=True)
     print("[bench] open a discussion on the dataset with these numbers and the mode (specialist or general).", flush=True)
     proc.terminate()

bench/submit_results.py

@@ -1,16 +1,20 @@
 """Create the Typed Decisions eval YAML and open a PR on the model repo.
 
 Usage:
-    python submit_results.py MODEL_ID [RESULTS_JSON] [--mode general|specialist] [--dry-run]
+    python submit_results.py MODEL_ID [RESULTS_JSON] [--mode general|specialist]
+                             [--predictions PATH] [--dry-run]
 
 Reads the summary produced by score_with_jobs.py (accuracy, kl_from_gold,
 brier, ece), writes the Hub eval-results YAML, and opens a pull request on
-the model repo (`.eval_results/<model>.yaml`). With --dry-run the YAML is
-printed instead of uploaded.
+the model repo (`.eval_results/<model>.yaml`). When predictions.jsonl from the
+same run is present (next to RESULTS_JSON unless --predictions says otherwise), the
+PR also adds it as `eval/typed-decisions/predictions.jsonl` so the scores can be
+re-derived. With --dry-run the YAML is printed instead of uploaded.
 """
 import argparse
 import datetime
 import json
+import os
 
 BENCH = "LocalLLaMA/typed-decisions"
 TASKS = ["accuracy", "kl_from_gold", "brier", "ece"]
@@ -45,6 +49,9 @@ def main():
     ap.add_argument("results", nargs="?", default="results.json", help="summary JSON from score_with_jobs.py")
     ap.add_argument("--mode", choices=["general", "specialist"], default="general",
                     help="specialist = fitted on the benchmark's train split")
+    ap.add_argument("--predictions", default=None,
+                    help="per-decision probabilities from score_with_jobs.py (default: predictions.jsonl next to "
+                         "the results file), added to the PR if present")
     ap.add_argument("--dry-run", action="store_true", help="print the YAML instead of opening a PR")
     args = ap.parse_args()
 
@@ -61,14 +68,18 @@ def main():
         print(yaml_text)
         return
 
-    from huggingface_hub import HfApi
+    from huggingface_hub import CommitOperationAdd, HfApi
 
+    ops = [CommitOperationAdd(path_in_repo=f".eval_results/{file_name}", path_or_fileobj=yaml_text.encode())]
+    predictions = args.predictions or os.path.join(os.path.dirname(args.results), "predictions.jsonl")
+    if os.path.exists(predictions):
+        ops.append(CommitOperationAdd(path_in_repo="eval/typed-decisions/predictions.jsonl",
+                                      path_or_fileobj=predictions))
     api = HfApi()
-    commit = api.upload_file(
-        path_or_fileobj=yaml_text.encode(),
-        path_in_repo=f".eval_results/{file_name}",
+    commit = api.create_commit(
         repo_id=args.model_id,
         repo_type="model",
+        operations=ops,
         create_pr=True,
         commit_message="Add Typed Decisions eval results",
         commit_description=(
@@ -78,7 +89,7 @@ def main():
         ),
     )
     pr_url = getattr(commit, "pr_url", None) or getattr(commit, "pull_request_url", "")
-    print(f"[submit] eval YAML opened as a PR on {args.model_id}: {pr_url}", flush=True)
+    print(f"[submit] eval results opened as a PR on {args.model_id}: {pr_url}", flush=True)
 
 
 if __name__ == "__main__":