filippogreppi commited on
Commit
bd71e6d
·
verified ·
1 Parent(s): f05201d

Add v5 and CrossNER results

Browse files
Files changed (1) hide show
  1. README.md +20 -11
README.md CHANGED
@@ -44,21 +44,30 @@ traces discuss the old output format.
44
  | `openvocab_train_v4_labelonly.jsonl` | 3,131 | Identical to v4 with an **empty** `<think>` block: the control that isolates the effect of the reasoning. |
45
  | `openvocab_train_v4_rejected.jsonl` | 3,131 | v4 recipe applied to CoTs **rejected** by the jury (same size as v4). |
46
  | `openvocab_train_v4_all.jsonl` | 10,901 | v4 recipe on **all** usable CoTs, without the jury (3,131 accepted + 7,770 rejected). |
 
 
47
 
48
  No file contains rows of the in-domain test set `universalmedner/pilener-eval-subset`.
49
 
50
- ## Main results (F1 type-agnostic, mean over the 9 biomedical benchmarks / Pile-NER test)
51
 
52
- | Model | Biomedical | Pile-NER |
53
- |---|---|---|
54
- | Qwen3-4B base (thinking) | **0.602** | 0.573 |
55
- | Qwen3-4B SFT v4 | 0.535 | 0.608 |
56
- | Qwen3-4B SFT v4 label-only | 0.500 | **0.645** |
57
- | Qwen3-1.7B base (thinking) | **0.557** | 0.492 |
58
- | Qwen3-1.7B SFT v4 | 0.487 | 0.567 |
59
- | Qwen3-1.7B SFT v4, 1 epoch | 0.510 | 0.550 |
60
- | Qwen3-1.7B SFT v4 rejected | 0.446 | 0.598 |
61
- | Qwen3-1.7B SFT v4 all | 0.484 | **0.621** |
 
 
 
 
 
 
 
62
 
63
  Training: LoRA r=16, alpha=32, lr 1e-4, 2 epochs (unless noted), loss on the
64
  assistant turn only.
 
44
  | `openvocab_train_v4_labelonly.jsonl` | 3,131 | Identical to v4 with an **empty** `<think>` block: the control that isolates the effect of the reasoning. |
45
  | `openvocab_train_v4_rejected.jsonl` | 3,131 | v4 recipe applied to CoTs **rejected** by the jury (same size as v4). |
46
  | `openvocab_train_v4_all.jsonl` | 10,901 | v4 recipe on **all** usable CoTs, without the jury (3,131 accepted + 7,770 rejected). |
47
+ | `openvocab_train_v5_teacher_qwen_3131.jsonl` | 3,131 | **v5**: v4 recipe, but the answer is the teacher's OWN entities (produced at the end of the same reasoning, as in ReasoningNER), so reasoning and answer never contradict each other; requested types = the teacher's labels in UPPER_SNAKE (e.g. `MEDICAL_CONDITION`); rows with a teacher entity not verbatim in the text are dropped. |
48
+ | `openvocab_train_v5_teacher_medgemma_3131.jsonl` | 3,131 | Same examples as v5 qwen, in MedGemma's reasoning format. |
49
 
50
  No file contains rows of the in-domain test set `universalmedner/pilener-eval-subset`.
51
 
52
+ ## Main results (F1 type-agnostic, mean over each test set)
53
 
54
+ | Model | Biomedical (9) | Pile-NER | CrossNER (5) |
55
+ |---|---|---|---|
56
+ | Qwen3-4B base (thinking) | **0.602** | 0.573 | **0.791** |
57
+ | Qwen3-4B SFT v4 | 0.535 | 0.608 | 0.751 |
58
+ | Qwen3-4B SFT v4 label-only | 0.500 | **0.645** | 0.716 |
59
+ | Qwen3-4B SFT v5 | 0.534 | 0.417 | 0.769 |
60
+ | MedGemma-1.5-4B base | 0.522 | 0.469 | 0.746 |
61
+ | MedGemma-1.5-4B SFT v5 | 0.527 | 0.436 | 0.783 |
62
+ | Qwen3-1.7B base (thinking) | 0.557 | 0.492 | – |
63
+ | Qwen3-1.7B SFT v4 | 0.487 | 0.567 | – |
64
+ | Qwen3-1.7B SFT v4, 1 epoch | 0.510 | 0.550 | – |
65
+ | Qwen3-1.7B SFT v4 rejected | 0.446 | 0.598 | – |
66
+ | Qwen3-1.7B SFT v4 all | 0.484 | 0.621 | – |
67
+
68
+ MedGemma v5 improves its base on 11 of 15 test sets, significantly on CrossNER (+0.037, 95% CI
69
+ [+0.027, +0.047]); on the biomedical mean it is on par (AnatEM drops, the teacher treats generic
70
+ anatomical nouns as non-entities). Adapters: `universalmedner/*-lora`.
71
 
72
  Training: LoRA r=16, alpha=32, lr 1e-4, 2 epochs (unless noted), loss on the
73
  assistant turn only.