ToMMeR-phi-4_L3_R64

ToMMeR is a lightweight probing model extracting emergent mention detection capabilities from early layers representations of any LLM backbone, achieving high Zero Shot recall across a wide set of 13 NER benchmarks.

Checkpoint Details

Property	Value
Base LLM	phi-4
Layer	3
#Params	660.5K

Usage

Installation

Our code can be installed with pip+git, Please visit the repository for more details.

pip install git+https://github.com/VictorMorand/llm2ner.git

Fancy Outputs

import llm2ner
from llm2ner import ToMMeR

tommer = ToMMeR.from_pretrained("llm2ner/ToMMeR-phi-4_L3_R64")
# load Backbone llm, optionnally cut the unused layer to save GPU space.
llm = llm2ner.utils.load_llm( tommer.llm_name, cut_to_layer=tommer.layer,) 
tommer.to(llm.device)

text = "Large language models are awesome. While trained on language modeling, they exhibit emergent Zero Shot abilities that make them suitable for a wide range of tasks, including Named Entity Recognition (NER). "

#fancy interactive output
outputs = llm2ner.plotting.demo_inference( text, tommer, llm,
    decoding_strategy="threshold",  # or "greedy" for flat segmentation
    threshold=0.5, # default 50%
    show_attn=True,
)

Large PRED language PRED models are awesome . While trained on language PRED modeling , they exhibit emergent PRED abilities that make them suitable for a wide range of tasks PRED , including Named PRED Entity Recognition ( NER PRED ) .

Raw inference

By default, ToMMeR outputs span probabilities, but we also propose built-in options for decoding entities.

• Inputs:
  • tokens (batch, seq): tokens to process,
  • model: LLM to extract representation from.
• Outputs: (batch, seq, seq) matrix (masked outside valid spans)

tommer = ToMMeR.from_pretrained("llm2ner/ToMMeR-phi-4_L3_R64")
# load Backbone llm, optionnally cut the unused layer to save GPU space.
llm = llm2ner.utils.load_llm( tommer.llm_name, cut_to_layer=tommer.layer,) 
tommer.to(llm.device)

#### Raw Inference
text = ["Large language models are awesome"]
print(f"Input text: {text[0]}")

#tokenize in shape (1, seq_len)
tokens = model.tokenizer(text, return_tensors="pt")["input_ids"].to(device)
# Output raw scores
output = tommer.forward(tokens, model) # (batch_size, seq_len, seq_len)
print(f"Raw Output shape: {output.shape}")

#use given decoding strategy to infer entities
entities = tommer.infer_entities(tokens=tokens, model=model, threshold=0.5, decoding_strategy="greedy")
str_entities = [ model.tokenizer.decode(tokens[0,b:e+1]) for b, e in entities[0]]
print(f"Predicted entities: {str_entities}")

>>> Input text: Large language models are awesome
>>> Raw Output shape: torch.Size([1, 6, 6])
>>> Predicted entities: ['Large language models']

Please visit the repository for more details and a demo notebook.

Evaluation Results

dataset	precision	recall	f1	n_samples
MultiNERD	0.2066	0.9561	0.3398	154144
CoNLL 2003	0.2633	0.7099	0.3842	16493
CrossNER_politics	0.2806	0.9431	0.4325	1389
CrossNER_AI	0.3175	0.9097	0.4708	879
CrossNER_literature	0.3272	0.8795	0.477	916
CrossNER_science	0.3236	0.9219	0.4791	1193
CrossNER_music	0.3523	0.9213	0.5097	945
ncbi	0.1134	0.847	0.2	3952
FabNER	0.3049	0.6933	0.4236	13681
WikiNeural	0.1966	0.9325	0.3247	92672
GENIA_NER	0.2133	0.9306	0.347	16563
ACE 2005	0.2414	0.3163	0.2738	8230
Ontonotes	0.2448	0.6719	0.3589	42193
Aggregated	0.2167	0.8764	0.3475	353250
Mean	0.2604	0.8179	0.3862	353250

Citation

If using this model or the approach, please cite the associated paper:

@misc{morand2025tommerefficiententity,
      title={ToMMeR -- Efficient Entity Mention Detection from Large Language Models}, 
      author={Victor Morand and Nadi Tomeh and Josiane Mothe and Benjamin Piwowarski},
      year={2025},
      eprint={2510.19410},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.19410}, 
}

License

Apache-2.0 (see repository for full text).