File size: 1,439 Bytes
a40d454
 
 
 
ba17454
a40d454
 
ba17454
c5bbe7a
ba17454
a40d454
 
ba17454
a40d454
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
---
license: other
language: [en]
tags: [ner, llm-as-a-judge, jury, chain-of-thought]
pretty_name: LLM-Jury Scores and Accepted Teacher Outputs
---

# LLM-Jury Scores and Accepted Teacher Outputs

Part of the code, data and models of the bachelor thesis *Structuring Universal Chain-of-thought for Named Entity Recognition* (Department of Computer Science and Engineering, University of Bologna, 2026). All the repositories cited by name belong to the same organization.

Three judge LLMs, from different model families, independently scored every
teacher extraction in `teacher-cot-pile-ner` against the
Pile-NER gold, on two axes:

- **completeness**: fraction of the gold spans meaningfully covered by the extraction;
- **correctness**: fraction of the extra, non-gold entities that are justified by the text.

| File | Content |
|---|---|
| `jury_scores_gemma-3-4b-it.jsonl` | scores of `google/gemma-3-4b-it` |
| `jury_scores_DeepSeek-R1-Distill-Llama-8B.jsonl` | scores of `deepseek-ai/DeepSeek-R1-Distill-Llama-8B` |
| `jury_scores_granite-4-1-8b.jsonl` | scores of `ibm-granite/granite-4.1-8b` |
| `jury_filtered_train.jsonl` | the **6,895** accepted rows (mean completeness ≥ 0.6 and mean correctness ≥ 0.6), in the original open-schema chat format used for the first MedGemma SFT |

Each score row has `id`, `judge_model`, `completeness_score`,
`completeness_notes`, `correctness_score`, `correctness_notes`, `raw_response`.