KomeijiForce/Cuckoo_C4_1M

Cuckoo 🐦 [Github]

Cuckoo: An IE Free Rider Hatched by Massive Nutrition in LLM's Nest is a small (300M) information extraction (IE) model that imitates the next token prediction paradigm of large language models. Instead of retrieving from the vocabulary, Cuckoo predicts the next tokens by tagging them in the given input context as shown below:

Cuckoo is substantially different from previous IE pre-training because it can use any text resource to enhance itself, especially by taking a free ride on data curated for LLMs!

Currently, we open-source checkpoints of Cuckoos that are pre-trained on:

1. 100M next tokens extraction (NTE) instances converted from C4. (Cuckoo-C4 🐦)

2. Cuckoo-C4 + 2.6M next token extraction (NTE) instances converted from a supervised fine-tuning dataset, TuluV3. (Cuckoo-C4-Instruct 🐦🛠️)

3. Cuckoo-C4-Instruct + MultiNERD, MetaIE, NuNER, MRQA (excluding SQuAD, DROP). (Cuckoo-C4-Rainbow 🌈🐦🛠️)

4. Cuckoo-C4-Rainbow + Multiple NER Datasets, WizardLM Dataset, Multiple Choice QA Datasets, MMLU, SQuAD, DROP, MNLI, SNLI. (Cuckoo-C4-Super-Rainbow 🦸🌈🐦🛠️)

• This dataset only contains the first 1M split for C4 NTE, you can convert the full C4 (and other corpus) with this script.

Performance Demonstration 🚀

Begin your journey with Cuckoo to experience unimaginable adaptation efficiency for all kinds of IE tasks!

	CoNLL2003	BioNLP2004	MIT-Restaurant	MIT-Movie	Avg.	CoNLL2004	ADE	Avg.	SQuAD	SQuAD-V2	DROP	Avg.
OPT-C4-TuluV3	50.24	39.76	58.91	56.33	50.56	47.14	45.66	46.40	39.80	53.81	31.00	41.54
RoBERTa	33.75	32.91	62.15	58.32	46.80	34.16	2.15	18.15	31.86	48.55	9.16	29.86
MRQA	72.45	55.93	68.68	66.26	65.83	66.23	67.44	66.84	80.07	66.22	54.46	66.92
MultiNERD	66.78	54.62	64.16	66.30	60.59	57.52	45.10	51.31	42.85	50.99	30.12	41.32
NuNER	74.15	56.36	68.57	64.88	65.99	65.12	63.71	64.42	61.60	52.67	37.37	50.55
MetaIE	71.33	55.63	70.08	65.23	65.57	64.81	64.40	64.61	74.59	62.54	30.73	55.95
Cuckoo 🐦🛠️	73.60	57.00	67.63	67.12	66.34	69.57	71.70	70.63	77.47	64.06	54.25	65.26
└─ Only Pre-train 🐦	72.46	55.87	66.87	67.23	65.61	68.14	69.39	68.77	75.64	63.36	52.81	63.94
└─ Only Post-train	72.80	56.10	66.02	67.10	65.51	68.66	69.75	69.21	77.05	62.39	54.80	64.75
Rainbow Cuckoo 🌈🐦🛠️	79.94	58.39	70.30	67.00	68.91	70.47	76.05	73.26	86.57	69.41	64.64	73.54

Quick Experience with Cuckoo in Next Tokens Extraction ⚡

We recommend using the strongest Super Rainbow Cuckoo 🦸🌈🐦🛠️ for zero-shot extraction.

1️⃣ First load the model and the tokenizers

from transformers import AutoModelForTokenClassification, AutoTokenizer
import torch
import spacy

nlp = spacy.load("en_core_web_sm")

device = torch.device("cuda:0")
path = f"KomeijiForce/Cuckoo-C4-Super-Rainbow"
tokenizer = AutoTokenizer.from_pretrained(path)
tagger = AutoModelForTokenClassification.from_pretrained(path).to(device)

2️⃣ Define the next tokens extraction function

def next_tokens_extraction(text):

    def find_sequences(lst):
        sequences = []
        i = 0
        while i < len(lst):
            if lst[i] == 0:
                start = i
                end = i
                i += 1
                while i < len(lst) and lst[i] == 1:
                    end = i
                    i += 1
                sequences.append((start, end+1))
            else:
                i += 1
        return sequences

    text = " ".join([token.text for token in nlp(text)])

    inputs = tokenizer(text, return_tensors="pt").to(device)
    tag_predictions = tagger(**inputs).logits[0].argmax(-1)

    predictions = [tokenizer.decode(inputs.input_ids[0, seq[0]:seq[1]]).strip() for seq in find_sequences(tag_predictions)]
    
    return predictions

3️⃣ Call the function for extraction!

Case 1: Basic entity and relation understanding

text = "Tom and Jack went to their trip in Paris."

for question in [
    "What is the person mentioned here?",
    "What is the city mentioned here?",
    "Who goes with Tom together?",
    "What do Tom and Jack go to Paris for?",
    "Where does George live in?",
]:
    prompt = f"User:\n\n{text}\n\nQuestion: {question}\n\nAssistant:"
    predictions = next_tokens_extraction(prompt)
    print(question, predictions)

You will get things like,

What is the person mentioned here? ['Tom', 'Jack']
What is the city mentioned here? ['Paris']
Who goes with Tom together? ['Jack']
What do Tom and Jack go to Paris for? ['trip']
Where does George live in? []

where [] indicates Cuckoo thinks there to be no next tokens for extraction.

Case 2: Longer context

passage = f'''Ludwig van Beethoven (17 December 1770 – 26 March 1827) was a German composer and pianist. He is one of the most revered figures in the history of Western music; his works rank among the most performed of the classical music repertoire and span the transition from the Classical period to the Romantic era in classical music. His early period, during which he forged his craft, is typically considered to have lasted until 1802. From 1802 to around 1812, his middle period showed an individual development from the styles of Joseph Haydn and Wolfgang Amadeus Mozart, and is sometimes characterised as heroic. During this time, Beethoven began to grow increasingly deaf. In his late period, from 1812 to 1827, he extended his innovations in musical form and expression.'''

for question in [
    "What are the people mentioned here?",
    "What is the job of Beethoven?",
    "How famous is Beethoven?",
    "When did Beethoven's middle period showed an individual development?",
]:
    text = f"User:\n\n{passage}\n\nQuestion: {question}\n\nAssistant:"
    predictions = next_tokens_extraction(text)
    print(question, predictions)

You will get things like,

What are the people mentioned here? ['Ludwig van Beethoven', 'Joseph Haydn', 'Wolfgang Amadeus Mozart']
What is the job of Beethoven? ['composer and pianist']
How famous is Beethoven? ['one of the most revered figures in the history of Western music']
When did Beethoven's middle period showed an individual development? ['1802']

Case 3: Knowledge quiz

for obj in ["grass", "sea", "fire", "night"]:
    text = f"User:\n\nChoices:\nred\nblue\ngreen.\n\nQuestion: What is the color of the {obj}?\n\nAssistant:\n\nAnswer:"
    predictions = next_tokens_extraction(text)
    print(obj, predictions)

You will get things like,

grass ['green']
sea ['blue']
fire ['red']
night []

which shows Cuckoo is not extracting any plausible spans but has the knowledge to understand the context.

Few-shot Adaptation 🎯

Cuckoo 🐦 is an expert in few-shot adaptation to your own tasks, taking CoNLL2003 as an example, run bash run_downstream.sh conll2003.5shot KomeijiForce/Cuckoo-C4-Rainbow, you will get a fine-tuned model in models/cuckoo-conll2003.5shot. Then you can benchmark the model with the script python eval_conll2003.py, which will show you an F1 performance of around 80.

You can also train the adaptation to machine reading comprehension (SQuAD), run bash run_downstream.sh squad.32shot KomeijiForce/Cuckoo-C4-Rainbow, you will get a fine-tuned model in models/cuckoo-squad.32shot. Then you can benchmark the model with the script python eval_squad.py, which will show you an F1 performance of around 88.

For fine-tuning your own task, you need to create a Jsonlines file, each line contains {"words": [...], "ner": [...]}, For example:

{"words": ["I", "am", "John", "Smith", ".", "Person", ":"], "ner": ["O", "O", "B", "I", "O", "O", "O"]}

which indicates "John Smith" to be predicted as the next tokens.

You can refer to some prompts shown below for beginning:

Type	User Input	Assistant Response
Entity	User: [Context] Question: What is the [Label] mentioned?	Assistant: Answer: The [Label] is
Relation (Kill)	User: [Context] Question: Who does [Entity] kill?	Assistant: Answer: [Entity] kills
Relation (Live)	User: [Context] Question: Where does [Entity] live in?	Assistant: Answer: [Entity] lives in
Relation (Work)	User: [Context] Question: Who does [Entity] work for?	Assistant: Answer: [Entity] works for
Relation (Located)	User: [Context] Question: Where is [Entity] located in?	Assistant: Answer: [Entity] is located in
Relation (Based)	User: [Context] Question: Where is [Entity] based in?	Assistant: Answer: [Entity] is based in
Relation (Adverse)	User: [Context] Question: What is the adverse effect of [Entity]?	Assistant: Answer: The adverse effect of [Entity] is
Query	User: [Context] Question: [Question]	Assistant: Answer:
Instruction (Entity)	User: [Context] Question: What is the [Label] mentioned? ([Instruction])	Assistant: Answer: The [Label] is
Instruction (Query)	User: [Context] Question: [Question] ([Instruction])	Assistant: Answer:

After building your own downstream dataset, save it into my_downstream.json, and then run the command bash run_downstream.sh my_downstream KomeijiForce/Cuckoo-C4-Rainbow. You will find an adapted Cuckoo in models/cuckoo-my_downstream.

Fly your own Cuckoo 🪽

We include the script to transform texts to NTE instances in the file nte_data_collection.py, which takes C4 as an example, the converted results can be checked in cuckoo.c4.example.json. The script is designed to be easily adapted to other resources like entity, query, and questions and you can modify your own data to NTE to fly your own Cuckoo! Run the run_cuckoo.sh script to try an example pre-training.

python run_ner.py \
  --model_name_or_path roberta-large \
  --train_file cuckoo.c4.example.json \
  --output_dir models/cuckoo-c4-example \
  --per_device_train_batch_size 4\
  --gradient_accumulation_steps 16\
  --num_train_epochs 1\
  --save_steps 1000\
  --learning_rate 0.00001\
  --do_train \
  --overwrite_output_dir

You will get an example Cuckoo model in models/cuckoo-c4-example, it might not perform well if you pre-train with too little data. You may adjust the hyperparameters inside nte_data_collection.py or modify the conversion for your own resources to enable better pre-training performance.

🐾 Citation

@article{DBLP:journals/corr/abs-2502-11275,
  author       = {Letian Peng and
                  Zilong Wang and
                  Feng Yao and
                  Jingbo Shang},
  title        = {Cuckoo: An {IE} Free Rider Hatched by Massive Nutrition in {LLM}'s Nest},
  journal      = {CoRR},
  volume       = {abs/2502.11275},
  year         = {2025},
  url          = {https://doi.org/10.48550/arXiv.2502.11275},
  doi          = {10.48550/arXiv.2502.11275},
  eprinttype   = {arXiv},
  eprint       = {2502.11275},
  timestamp    = {Mon, 17 Feb 2025 19:32:20 +0000},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2502-11275.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}