SRK-CobuildX commited on
Commit
d2a8d8f
·
verified ·
1 Parent(s): 0dff23d

CallOpeningDetX v0.1

Browse files
Files changed (3) hide show
  1. README.md +7 -4
  2. checksums.sha256 +2 -2
  3. stats.json +232 -0
README.md CHANGED
@@ -286,7 +286,8 @@ All-fields exact per test subset (95% interval in brackets):
286
  |---|---|---|---|
287
  | Keyword rules | 0.977 [0.965–0.988] | 0.665 [0.600–0.730] | 0.495 [0.462–0.526] |
288
  | Qwen2.5-7B-Instruct, zero-shot | 0.960 [0.943–0.975] | 0.930 [0.895–0.965] | 0.425 [0.395–0.457] |
289
- | CobuildX call opening checker 17M (ours) | 0.988 [0.980–0.997] | 0.970 [0.945–0.990] | 0.862 [0.840–0.883] |
 
290
 
291
  Challenge subset, per field:
292
 
@@ -294,13 +295,15 @@ Challenge subset, per field:
294
  |---|---|---|---|---|---|---|---|
295
  | Keyword rules | 0.742 | 0.733 | 0.851 / 0.847 | 0.799 | 0.975 | 0.617 / 0.345 | 0.820 / 0.826 |
296
  | Qwen2.5-7B-Instruct, zero-shot | 0.750 | 0.723 | 0.801 / 0.528 | 0.766 | 0.945 | 0.706 / 0.321 | 0.803 / 0.788 |
297
- | CobuildX call opening checker 17M (ours) | 0.981 | 0.994 | 0.963 / 0.961 | 0.854 | 1.000 | 0.959 / 0.862 | 0.993 / 0.983 |
 
298
 
299
  - **Keyword rules**: hand-written patterns for names, institution greetings, notice phrases, purposes and negations, applied to the opening turns.
300
  - **Qwen2.5-7B-Instruct, zero-shot**: the annotation guide as system prompt, one call per request, greedy decoding; evidence quotes matched to the text for spans. Replies that were not valid JSON count as wrong.
301
- - **[CobuildX call opening checker 17M](https://huggingface.co/CobuildX-AI/cbx-bfs-call-opening-checker-17M)** (ours): a fine-tuned Ettin encoder (17M parameters) with field heads and an evidence-span tagger.
 
302
 
303
- Nothing was tuned on test. **The CobuildX model is ours.** Its training data was made with the same construction
304
  method as the challenge subset (rewritten openings, invented names), although the wording families and name pools
305
  of the test set were held out, and its checkpoint was chosen on dev. That may give it a home advantage; please read
306
  its numbers with this in mind.
 
286
  |---|---|---|---|
287
  | Keyword rules | 0.977 [0.965–0.988] | 0.665 [0.600–0.730] | 0.495 [0.462–0.526] |
288
  | Qwen2.5-7B-Instruct, zero-shot | 0.960 [0.943–0.975] | 0.930 [0.895–0.965] | 0.425 [0.395–0.457] |
289
+ | Jev (TypeSafe), zero-shot | 0.997 [0.992–1.000] | 0.995 [0.985–1.000] | 0.883 [0.863–0.902] |
290
+ | `cbx-bfs-call-opening-checker-17M` (ours) | 0.988 [0.980–0.997] | 0.970 [0.945–0.990] | 0.862 [0.840–0.883] |
291
 
292
  Challenge subset, per field:
293
 
 
295
  |---|---|---|---|---|---|---|---|
296
  | Keyword rules | 0.742 | 0.733 | 0.851 / 0.847 | 0.799 | 0.975 | 0.617 / 0.345 | 0.820 / 0.826 |
297
  | Qwen2.5-7B-Instruct, zero-shot | 0.750 | 0.723 | 0.801 / 0.528 | 0.766 | 0.945 | 0.706 / 0.321 | 0.803 / 0.788 |
298
+ | Jev (TypeSafe), zero-shot | 0.929 | 0.975 | 0.974 / 0.961 | 0.980 | 0.995 | — | 0.947 / 0.947 |
299
+ | `cbx-bfs-call-opening-checker-17M` (ours) | 0.981 | 0.994 | 0.963 / 0.961 | 0.854 | 1.000 | 0.959 / 0.862 | 0.993 / 0.983 |
300
 
301
  - **Keyword rules**: hand-written patterns for names, institution greetings, notice phrases, purposes and negations, applied to the opening turns.
302
  - **Qwen2.5-7B-Instruct, zero-shot**: the annotation guide as system prompt, one call per request, greedy decoding; evidence quotes matched to the text for spans. Replies that were not valid JSON count as wrong.
303
+ - **Jev (TypeSafe), zero-shot**: TypeSafe's decision model through its System One API, five choice questions per call (one per field; purpose as one choice over the purpose sets), each value with a one-line criterion written from the annotation guide. It returns no evidence, so span scores do not apply.
304
+ - **[`cbx-bfs-call-opening-checker-17M`](https://huggingface.co/CobuildX-AI/cbx-bfs-call-opening-checker-17M)** (ours, CobuildX AI): a fine-tuned Ettin encoder (17M parameters) with field heads and an evidence-span tagger.
305
 
306
+ Nothing was tuned on test. **`cbx-bfs-call-opening-checker-17M` is ours** (built by the authors of this benchmark). Its training data was made with the same construction
307
  method as the challenge subset (rewritten openings, invented names), although the wording families and name pools
308
  of the test set were held out, and its checkpoint was chosen on dev. That may give it a home advantage; please read
309
  its numbers with this in mind.
checksums.sha256 CHANGED
@@ -1,9 +1,9 @@
1
  bf3374919cd5c314782b355a06ae2d0081da132518fa74eadf3c9d309d11606a DATASHEET.md
2
  9d242f2775d7a1d61249e49ed08bfcca3d8759782ad4bfdaa6dcfff830765054 LICENSE
3
- 658f67903d3092d423e003c9b7dd1b54566ae0dd75d7687731aac76cae86faea README.md
4
  c3df34f4b9d8030f5ac9c76be6df4a8a7ba6552c1490312f23acf1fef9f1ea9a annotation_guide.md
5
  0705935f7943055d16453df873360e85c9261e0813c360b1bc8f346c322add6e data/dev.jsonl
6
  ed7d9f2a529d0dcefbe016ff26a91c2a7f89e74a5fedd52de7299529932f22c6 data/test.jsonl
7
  4a5b396b5012238d71ce5becc63833b5cf86df1f2bed10787916f800dd4c576f render.py
8
  026dd0394b2c6920b0ca8a0ee3dc8264d0d472bfeb24caa4278e8380e260d88d score.py
9
- b6a735d57fb5db227ebadaf7f7dc05394eeba3a7a726ebe437fc7061111e4b0f stats.json
 
1
  bf3374919cd5c314782b355a06ae2d0081da132518fa74eadf3c9d309d11606a DATASHEET.md
2
  9d242f2775d7a1d61249e49ed08bfcca3d8759782ad4bfdaa6dcfff830765054 LICENSE
3
+ 1444c72f9c48ae8b14df6111439162b654a344e720d50431fae2cf42bc3814f3 README.md
4
  c3df34f4b9d8030f5ac9c76be6df4a8a7ba6552c1490312f23acf1fef9f1ea9a annotation_guide.md
5
  0705935f7943055d16453df873360e85c9261e0813c360b1bc8f346c322add6e data/dev.jsonl
6
  ed7d9f2a529d0dcefbe016ff26a91c2a7f89e74a5fedd52de7299529932f22c6 data/test.jsonl
7
  4a5b396b5012238d71ce5becc63833b5cf86df1f2bed10787916f800dd4c576f render.py
8
  026dd0394b2c6920b0ca8a0ee3dc8264d0d472bfeb24caa4278e8380e260d88d score.py
9
+ a6cbe1e10fee92831a4520ba1899daf093e1dd8c7c99a6f3f33cdc701c7f3317 stats.json
stats.json CHANGED
@@ -581,6 +581,238 @@
581
  }
582
  }
583
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
584
  "cbx-bfs-call-opening-checker-17M": {
585
  "missing_predictions": 0,
586
  "all": {
 
581
  }
582
  }
583
  },
584
+ "jev-zeroshot": {
585
+ "missing_predictions": 0,
586
+ "all": {
587
+ "n": 1750,
588
+ "agent_identification_acc": 0.9611,
589
+ "agent_identification_macro_f1": 0.9145,
590
+ "institution_identification_acc": 0.9857,
591
+ "institution_identification_macro_f1": 0.9498,
592
+ "recording_notice_acc": 0.9857,
593
+ "recording_notice_macro_f1": 0.9641,
594
+ "opening_coverage_acc": 0.9971,
595
+ "opening_coverage_macro_f1": 0.9809,
596
+ "field_macro_f1_mean": 0.9524,
597
+ "all_fields_exact": 0.9349,
598
+ "purpose_micro_f1": 0.9783,
599
+ "purpose_macro_f1": 0.9807,
600
+ "purpose_rows": 451,
601
+ "span_rows": 1750,
602
+ "span_f1_exact": 0.0,
603
+ "span_f1_overlap": 0.0,
604
+ "span_f1_overlap_by_type": {
605
+ "AGENT": 0.0,
606
+ "DISTRACTOR": 0.0,
607
+ "INST": 0.0,
608
+ "NEGATED": 0.0,
609
+ "NOTICE": 0.0,
610
+ "PURPOSE": 0.0
611
+ },
612
+ "policy_A_agree": 0.9714,
613
+ "policy_B_agree": 0.9714,
614
+ "all_fields_exact_ci95": [
615
+ 0.9229,
616
+ 0.9469
617
+ ],
618
+ "all_fields_exact_by_construction": {
619
+ "automated_notice_metadata": {
620
+ "n": 38,
621
+ "all_fields_exact": 0.8684
622
+ },
623
+ "automated_notice_turn": {
624
+ "n": 68,
625
+ "all_fields_exact": 0.8676
626
+ },
627
+ "cut_transcript": {
628
+ "n": 79,
629
+ "all_fields_exact": 0.7722
630
+ },
631
+ "distractor_agent": {
632
+ "n": 140,
633
+ "all_fields_exact": 0.8929
634
+ },
635
+ "distractor_customer": {
636
+ "n": 127,
637
+ "all_fields_exact": 0.8976
638
+ },
639
+ "invented_names": {
640
+ "n": 950,
641
+ "all_fields_exact": 0.8832
642
+ },
643
+ "late_mention": {
644
+ "n": 19,
645
+ "all_fields_exact": 0.6316
646
+ },
647
+ "negated_notice": {
648
+ "n": 78,
649
+ "all_fields_exact": 0.8333
650
+ },
651
+ "no_call_start_metadata": {
652
+ "n": 98,
653
+ "all_fields_exact": 0.8469
654
+ },
655
+ "notice_after_reason": {
656
+ "n": 192,
657
+ "all_fields_exact": 0.8333
658
+ },
659
+ "notice_in_first_turn": {
660
+ "n": 455,
661
+ "all_fields_exact": 0.8769
662
+ },
663
+ "opening_rewritten": {
664
+ "n": 753,
665
+ "all_fields_exact": 0.8645
666
+ }
667
+ }
668
+ },
669
+ "natural": {
670
+ "n": 600,
671
+ "agent_identification_acc": 1.0,
672
+ "agent_identification_macro_f1": 1.0,
673
+ "institution_identification_acc": 0.9983,
674
+ "institution_identification_macro_f1": 0.944,
675
+ "recording_notice_acc": 1.0,
676
+ "recording_notice_macro_f1": 1.0,
677
+ "opening_coverage_acc": 1.0,
678
+ "opening_coverage_macro_f1": 1.0,
679
+ "field_macro_f1_mean": 0.986,
680
+ "all_fields_exact": 0.9967,
681
+ "purpose_micro_f1": 0.875,
682
+ "purpose_macro_f1": 0.7,
683
+ "purpose_rows": 7,
684
+ "span_rows": 600,
685
+ "span_f1_exact": 0.0,
686
+ "span_f1_overlap": 0.0,
687
+ "span_f1_overlap_by_type": {
688
+ "AGENT": 0.0,
689
+ "INST": 0.0,
690
+ "NOTICE": 0.0,
691
+ "PURPOSE": 0.0
692
+ },
693
+ "policy_A_agree": 1.0,
694
+ "policy_B_agree": 1.0,
695
+ "all_fields_exact_ci95": [
696
+ 0.9917,
697
+ 1.0
698
+ ],
699
+ "all_fields_exact_by_construction": {}
700
+ },
701
+ "insurance": {
702
+ "n": 200,
703
+ "agent_identification_acc": 0.995,
704
+ "agent_identification_macro_f1": 0.9963,
705
+ "institution_identification_acc": 1.0,
706
+ "institution_identification_macro_f1": 1.0,
707
+ "recording_notice_acc": 1.0,
708
+ "recording_notice_macro_f1": 1.0,
709
+ "opening_coverage_acc": 1.0,
710
+ "opening_coverage_macro_f1": 1.0,
711
+ "field_macro_f1_mean": 0.9991,
712
+ "all_fields_exact": 0.995,
713
+ "purpose_micro_f1": null,
714
+ "purpose_macro_f1": null,
715
+ "purpose_rows": 0,
716
+ "span_rows": 200,
717
+ "span_f1_exact": 0.0,
718
+ "span_f1_overlap": 0.0,
719
+ "span_f1_overlap_by_type": {
720
+ "AGENT": 0.0,
721
+ "DISTRACTOR": 0.0,
722
+ "INST": 0.0
723
+ },
724
+ "policy_A_agree": 1.0,
725
+ "policy_B_agree": 1.0,
726
+ "all_fields_exact_ci95": [
727
+ 0.985,
728
+ 1.0
729
+ ],
730
+ "all_fields_exact_by_construction": {}
731
+ },
732
+ "challenge": {
733
+ "n": 950,
734
+ "agent_identification_acc": 0.9295,
735
+ "agent_identification_macro_f1": 0.8981,
736
+ "institution_identification_acc": 0.9747,
737
+ "institution_identification_macro_f1": 0.9478,
738
+ "recording_notice_acc": 0.9737,
739
+ "recording_notice_macro_f1": 0.9611,
740
+ "opening_coverage_acc": 0.9947,
741
+ "opening_coverage_macro_f1": 0.9809,
742
+ "field_macro_f1_mean": 0.947,
743
+ "all_fields_exact": 0.8832,
744
+ "purpose_micro_f1": 0.9799,
745
+ "purpose_macro_f1": 0.9823,
746
+ "purpose_rows": 444,
747
+ "span_rows": 950,
748
+ "span_f1_exact": 0.0,
749
+ "span_f1_overlap": 0.0,
750
+ "span_f1_overlap_by_type": {
751
+ "AGENT": 0.0,
752
+ "DISTRACTOR": 0.0,
753
+ "INST": 0.0,
754
+ "NEGATED": 0.0,
755
+ "NOTICE": 0.0,
756
+ "PURPOSE": 0.0
757
+ },
758
+ "policy_A_agree": 0.9474,
759
+ "policy_B_agree": 0.9474,
760
+ "all_fields_exact_ci95": [
761
+ 0.8632,
762
+ 0.9021
763
+ ],
764
+ "all_fields_exact_by_construction": {
765
+ "automated_notice_metadata": {
766
+ "n": 38,
767
+ "all_fields_exact": 0.8684
768
+ },
769
+ "automated_notice_turn": {
770
+ "n": 68,
771
+ "all_fields_exact": 0.8676
772
+ },
773
+ "cut_transcript": {
774
+ "n": 79,
775
+ "all_fields_exact": 0.7722
776
+ },
777
+ "distractor_agent": {
778
+ "n": 140,
779
+ "all_fields_exact": 0.8929
780
+ },
781
+ "distractor_customer": {
782
+ "n": 127,
783
+ "all_fields_exact": 0.8976
784
+ },
785
+ "invented_names": {
786
+ "n": 950,
787
+ "all_fields_exact": 0.8832
788
+ },
789
+ "late_mention": {
790
+ "n": 19,
791
+ "all_fields_exact": 0.6316
792
+ },
793
+ "negated_notice": {
794
+ "n": 78,
795
+ "all_fields_exact": 0.8333
796
+ },
797
+ "no_call_start_metadata": {
798
+ "n": 98,
799
+ "all_fields_exact": 0.8469
800
+ },
801
+ "notice_after_reason": {
802
+ "n": 192,
803
+ "all_fields_exact": 0.8333
804
+ },
805
+ "notice_in_first_turn": {
806
+ "n": 455,
807
+ "all_fields_exact": 0.8769
808
+ },
809
+ "opening_rewritten": {
810
+ "n": 753,
811
+ "all_fields_exact": 0.8645
812
+ }
813
+ }
814
+ }
815
+ },
816
  "cbx-bfs-call-opening-checker-17M": {
817
  "missing_predictions": 0,
818
  "all": {