Dataset: tner/ontonotes5 raw JSON shards
Train examples: 12000
Eval examples: 512
Fixed scored eval slice: first 64 validation examples after deterministic filtering/shuffle with seed 42
Negative ratio: 0.15

Training recipe

Base model: Qwen/Qwen3.5-0.8B
4-bit base + LoRA SFT
LoRA target modules: auto-discovered
LoRA rank/alpha/dropout: 16 / 32 / 0.05
Batch size: 2
Grad accumulation: 8
Max steps: 300
Learning rate: 2e-4
Max length: 384
Max new tokens: 96
Seed: 42

Metrics

Original logged winning run (Run 4):

entity_f1_pct: 89.5105
json_exact_match_pct: 81.2500
parseable_pct: 100.0000

Republished checkpoint pushed to this repo from the same recipe/commit with Hub push enabled:

entity_f1_pct: 94.4444
json_exact_match_pct: 92.1875
parseable_pct: 100.0000

Usage

python

from transformers import AutoTokenizer, AutoModelForImageTextToText
from peft import PeftModel

base_model = "Qwen/Qwen3.5-0.8B"
adapter_id = "Mike0021/qwen35-0.8b-ner-json-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(base_model, trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter_id)

Prompt format used during training:

system: Extract named entities from the sentence. Return strict JSON only with keys people, places, dates. Each value must be an array of strings. Use [] when empty.
user: Sentence: <sentence>

The model should answer with JSON only.