ICE v2 / LSREP MicroNER

This releases the retained original MicroNER checkpoint loaded by frozen ICE v2, evaluated in the LSREP paper. It is not a release of current ICE v3 research models.

Repository: Deepnar/ice.
Evaluated tag: v2-paper-eval.
Evaluated commit: 0521df9171b4a7d69f82d12d70497138c77b2678.
Original path: models/ner/ner_model.pt.
Original file size: 234,093 bytes; bare PyTorch state dict.
SHA-256: 23e596654065bde16f822db4b6dacf1830f67ae3613234bb7d808f5ea214c1b2.

The .pt file is copied byte for byte, without retraining or re-export. The frozen Git tag records the loader and checkpoint path, but does not contain the checkpoint blob or a historical checksum. The release hashes the retained original.

Architecture, labels and input

Linear(384,128) -> ReLU -> Dropout(0.2) -> Linear(128,64) -> ReLU -> Dropout(0.2) -> Linear(64,3); 57,731 trainable parameters. The exact frozen ner_model.py is included. Input shape is (batch,tokens,384); output is (batch,tokens,3). Disable dropout with eval(). The label order is:

0 = B-ENT   beginning of a generic entity
1 = I-ENT   continuation of a generic entity
2 = O       outside

These are generic entity spans, not typed person/organization/location classes. Each token is decoded by argmax. There is no CRF, contextual token encoder, softmax threshold, or typed-entity classification head in this artifact.

Required preprocessing and span reconstruction

Use the tokenizer from Qwen/Qwen3-Embedding-0.6B, revision 97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3. Tokenize the supplied text with return_offsets_mapping=True and add_special_tokens=False, then call convert_ids_to_tokens(). Pass those raw tokenizer token strings, including any subword markers, individually to SentenceTransformer.encode() as a list of separate inputs. Do not replace them with decoded whole words or transformer hidden states.

The frozen base has native width 1024; set truncate_dim=384. Preserve the snapshot's built-in pooling/normalization. Do not request additional normalize_embeddings=True, renormalize the 384-coordinate prefix, or use current ICE v3's native-width encoder path.

Use tokenizer character offsets to reconstruct spans. An I-ENT without an open B-ENT is ignored. Whitespace-adjacent predicted spans are joined, matching the frozen loader. Retrieval defaults to full text (max_chars=None); callers may explicitly truncate text before tokenization, as clustering does. The original frozen_ner_utils.py is supplied verbatim for audit; the included microner_inference.py implements its trained-model path without its regex fallback.

Minimal inference

The frozen tag uses Python 3.11, torch==2.11.0, sentence-transformers==5.5.1, and transformers==5.9.0.

import hashlib
import sys
from pathlib import Path
import torch
from huggingface_hub import snapshot_download
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer

folder = Path(snapshot_download("Deepnar/ice-v2-microner"))
sys.path.insert(0, str(folder))
from ner_model import MicroNER
from microner_inference import extract_entities

checkpoint = folder / "ner_model.pt"
assert hashlib.sha256(checkpoint.read_bytes()).hexdigest() == (
    "23e596654065bde16f822db4b6dacf1830f67ae3613234bb7d808f5ea214c1b2"
)
head = MicroNER()
head.load_state_dict(torch.load(checkpoint, map_location="cpu", weights_only=True))
head.eval()
base = "Qwen/Qwen3-Embedding-0.6B"
revision = "97b0c614be4d77ee51c0cef4e5f07c00f9eb65b3"
tokenizer = AutoTokenizer.from_pretrained(base, revision=revision)
embedder = SentenceTransformer(base, revision=revision, device="cpu", truncate_dim=384)
print(extract_entities(head, tokenizer, embedder, "Alice works at Example Labs in Berlin."))

For durable reuse, pin snapshot_download(..., revision=<release commit>) to the upload commit in the GitHub release verification report. This custom head is not a Transformers token-classification pipeline. Neither a successful load nor a nonempty span is a validation of entity correctness or graph utility.

Limitations and license

The paper establishes that the trained ICE v2 checkpoint was present and loaded; it does not establish task-specific NER recall or graph accuracy. No new quality measurement is claimed here. Independent token-string embeddings, generic BIO labels, subword segmentation and whitespace joining limit span accuracy. Private corpora, probes and training data are excluded; the release supports checkpoint/inference reuse, not independent regeneration of private training.

The ICE head and accompanying ICE code use the repository's Apache-2.0 license; LICENSE and NOTICE are included. The separately fetched Qwen base model is Apache-2.0 according to its pinned upstream model card. No upstream base weights, private text, caches, secrets, unrelated checkpoints or ICE v3 artifacts are uploaded.