structured-cot-ner/openvocab-sft

Type-Conditioned SFT Data (First Phase)

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.

Training sets for fine-tuning Qwen3 (thinking mode) on open-vocabulary NER: the entity types to extract are given in the prompt, and the target is the original Pile-NER annotation. Built from Pile-NER rows that have a teacher Chain-of-Thought (Qwen3-32B, open-schema), filtered by a jury of three LLM judges (see jury).

Format

One JSON object per line: {"id", "messages", "meta"}.

• system + user: exactly the prompt the baseline receives at evaluation time (same text as bioner-eval-subset), with the list of ENTITY TYPES to extract.
• assistant: <think>\n{teacher reasoning}\n</think>\n\n{JSON}, where the JSON has one key per requested type and the Pile-NER gold spans (de-duplicated per type) as values; absent types are answered with [].
• meta: agreement between the teacher reasoning and the gold that now follows it (n_gold_spans, gold_in_teacher_entities, gold_rejected_by_teacher, gold_mentioned_in_reasoning, teacher_entities_not_in_gold, reasoning_mentions_old_format, reasoning_source).

The teacher reasoning was generated for the open-schema task, so it does not always agree with the gold: on the 6,895 jury-accepted rows, the teacher explicitly rejects a gold entity in 41.5% of the examples and 30.4% of the traces discuss the old output format.

Files (one change at a time)

File	Rows	What changes
openvocab_train_v3_qwen.jsonl	3,131	Jury-accepted rows; removes examples whose reasoning rejects a gold entity or discusses the old output format; each prompt adds 2 absent "distractor" types (answered []); a few examples request no present type.
openvocab_train_v3_medgemma.jsonl	3,131	Same examples as v3, in MedGemma's reasoning format (<unused94>thought … <unused95>).
openvocab_train_v4_qwen.jsonl	3,131	v3, but each example requests only a random non-empty subset of the types present (with v3 the model put every entity it found under whatever type was requested, e.g. genes as DISEASE on NCBI).
openvocab_train_v4_labelonly.jsonl	3,131	Identical to v4 with an empty <think> block: the control that isolates the effect of the reasoning.
openvocab_train_v4_rejected.jsonl	3,131	v4 recipe applied to CoTs rejected by the jury (same size as v4).
openvocab_train_v4_all.jsonl	10,901	v4 recipe on all usable CoTs, without the jury (3,131 accepted + 7,770 rejected).
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.
openvocab_train_v5_teacher_medgemma_3131.jsonl	3,131	Same examples as v5 qwen, in MedGemma's reasoning format.
openvocab_train_v5_teacher_qwen_A_strict08.jsonl	427	v5 recipe on the outputs accepted with the jury threshold raised from 0.6 to 0.8: the stricter-jury ablation (Qwen3-1.7B).
openvocab_train_v5_teacher_qwen_B_random.jsonl	427	427 examples drawn at random from the same pool of v5 examples accepted at 0.6: the control of the stricter-jury ablation.

No file contains rows of the in-domain test set pilener-eval-subset.

Main results

Exact typed F1 (the metric of the thesis), mean over each group of test sets: a predicted mention is correct only if it equals a reference mention of the same type, ignoring case; only the requested types are scored. Every run is in eval-results. In the thesis, the versions v3, v4 and v5 of these files are called v1, v2 and v3. – = not evaluated.

Model	Biomedical (9)	Pile-NER	CrossNER (5)
Qwen3-4B base (thinking)	0.505	0.462	0.660
Qwen3-4B SFT v4	0.435	0.486	0.570
Qwen3-4B SFT v4 label-only	0.398	0.524	0.531
Qwen3-4B SFT v5	0.445	0.359	0.601
MedGemma-1.5-4B base	0.422	0.360	0.598
MedGemma-1.5-4B SFT v5	0.426	0.364	0.627
Qwen3-1.7B base (thinking)	0.461	0.384	0.604
Qwen3-1.7B SFT v3	0.395	–	0.522
Qwen3-1.7B SFT v4	0.394	0.450	0.565
Qwen3-1.7B SFT v4 label-only	0.392	0.489	0.515
Qwen3-1.7B SFT v4, 1 epoch	0.399	0.431	0.574
Qwen3-1.7B SFT v4 rejected	0.382	0.472	0.445
Qwen3-1.7B SFT v4 all	0.373	0.505	0.437
Qwen3-1.7B SFT v5	0.389	–	0.596
Qwen3-1.7B SFT v5, threshold 0.8 (A)	0.368	–	0.595
Qwen3-1.7B SFT v5, random 427 (B)	0.378	–	0.588

No Qwen3 student outperforms its base model on the biomedical benchmarks or on CrossNER; MedGemma v5 improves its base on CrossNER (0.627 against 0.598). Adapters: the *-lora adapters of this organization.

Training: LoRA r=16, alpha=32, lr 1e-4, 2 epochs (unless noted), loss on the assistant turn only.

Derived from Pile-NER: check its license before reuse beyond this project.