File size: 2,299 Bytes
947f4d8
 
 
4926e1b
947f4d8
 
4926e1b
947f4d8
796c741
947f4d8
 
796c741
 
947f4d8
 
 
 
 
 
 
796c741
 
 
 
947f4d8
796c741
 
 
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
29
---
tags: [named-entity-recognition, chain-of-thought, sft]
language: [en]
pretty_name: Training Files of the Free vs Structured Reasoning Students
---

# Training Files of the Free vs Structured Reasoning Students

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. Code: [GreppiFilippo/structured-cot-ner](https://github.com/GreppiFilippo/structured-cot-ner).

The supervised fine-tuning files of the second-phase students (the `*-cot-lora` adapters of the organization), built
from the train split of `pilener-teacher-cot-agreement-v1` (3,822 queries). Every file has the same queries, in the
same order, and the same target answer (the teacher's `answer`); only the reasoning differs.

Each row: `id` (the query id of the source dataset) and `messages`, a user turn (the prompt) and an assistant turn
`<think>\n{reasoning}</think>\n\n{answer}`, where the answer is a JSON object mapping every requested type to its
mentions. No system prompt; the loss is computed on the assistant turn only.

| File | Reasoning | Prompt | Built with |
|---|---|---|---|
| `pilener_paired_free_cot.jsonl` | the teacher's free reasoning (`free_cot`) | `free_prompt` | `05_openvocab_sft/build_paired_cot_dataset.py` |
| `pilener_paired_structured_cot.jsonl` | the teacher's structured reasoning (`structured_cot`): `TYPES`, `MENTIONS`, `EDGE` | `structured_prompt` | `05_openvocab_sft/build_paired_cot_dataset.py` |
| `pilener_paired_noedge_cot.jsonl` | the structured reasoning without the `EDGE` line | `structured_prompt` without the `EDGE` instruction | `05_openvocab_sft/make_noedge_data.py` |
| `pilener_paired_noedge_inv_cot.jsonl` | the structured reasoning without `EDGE`, with `MENTIONS` before `TYPES` | the same edit, with the two instructions swapped | `05_openvocab_sft/make_noedge_data.py --invert` |

The two variants without `EDGE` are derived from the existing structured traces (the line is removed, and in the
second the remaining fields are swapped), not regenerated by the teacher; they are the variants of the structured
format studied in the ablations of the thesis.