cstr/UltraBERT-Embed-150M-CrispEmbed-GGUF

UltraBERT-Embed-150M Q4_K GGUF (CrispEmbed, EXPERIMENTAL)

A quantized GGUF of dmis-lab/UltraBERT-Embed-150M by Jungwoo Park, Taewhoo Lee and Jaewoo Kang (DMIS Lab), for CrispEmbed's default-off native UltraBERT text-embedding path. All credit for the model belongs to the original authors.

Status: experimental. It clears CrispEmbed's Q4_K precision bar (final-embedding cosine > 0.94) and retrieval quality is close to the original, but it fails the stricter F32 vector and top-1 agreement gates (below).

Do not mix Q4_K vectors into an index built with F32 (or the original model) vectors. Per-row cosine to the original can drop to 0.9779 and top-1 agreement is 0.973; re-embed the whole corpus with the same file you query with.

Quality (SciFact, 300 queries over 5,483 rows, vs the original F32 PyTorch model)

Precision	Size	Min embedding cosine	Mean embedding cosine	nDCG@10	MRR@10	Recall@100	Top-1 agreement	Strict F32 vector gate	Status
Original (F32)	581.2 MiB	1	1	0.7667	0.7290	0.9750	reference	reference	reference
Q4_K q4k-last-four-ffn (this file)	550.3 MiB	0.9779	not measured	0.7649 (-0.0018)	0.7279 (-0.0011)	0.9767 (+0.0017)	0.973	fail (max abs normalized 0.0282)	experimental

• Measured by CrispEmbed run 40871150002 (runtime commit 018cf85b7bb6, portable build, hosted x86-64 CPU) on exactly these bytes. The protocol's own strict gates (F32 vector budget, top-1 >= 0.99, task degradation <= 0.001) fail and stay recorded; publication follows the owner's precision policy.
• Not measured: GPU backends, ARM/Apple, other corpora/languages, long-context behaviour, latency.

Files

File	Precision	Bytes	SHA256
ultrabert-embed-150m-q4_k-last-four-ffn.gguf	Q4_K-LAST-FOUR-FFN	576,997,440	087c9be0fa50ba57db74e650889521c985d306863c94aabf731764538eb4533c

• crispembed-quantize q4_k with token_embd and every attention matrix F32 and blk.0-19 FFN gate/up/down F32: only the last four layers' FFN matrices (12 tensors) are Q4_K, 206 tensors stay F32. Legacy UltraBERT header written by the runtime-018cf85b quantizer. Stored types: F32 206, Q4_K 12.

License and attribution

• The original model is released under the Apache License 2.0; this derived file is distributed under the same license. The model repository declares license: apache-2.0 but ships no LICENSE or NOTICE file, so LICENSE is the verbatim canonical Apache-2.0 text (SHA256 cfc7749b96f63bd31c3c42b5c471bf756814053e847c10f3eb003417bc523d30) and no NOTICE is included.
• UPSTREAM_README.md is the unmodified upstream model card at revision 367d5343b5a0164d57b810cedf5de4b8983d5984 (SHA256 27573272049ca66a4504b7e458887848b0fa0b7a5c6c68edcd774c6ce5070608).

Statement of changes (Apache-2.0 section 4(b))

Modified from the original by CrispEmbed: the model.safetensors weights of revision 367d5343b5a0164d57b810cedf5de4b8983d5984 were format-converted to GGUF (models/convert-ultrabert-to-gguf.py, F32, tokenizer metadata embedded; F32 SHA256 89478b9cbbad2fd1121be4031d760f1529c989fedf23b6b19251ba2c87422566) and then quantized with CrispEmbed's crispembed-quantize (recipe above). No weights were trained, fine-tuned, merged or otherwise altered.

How the file was produced

1. Converted and quantized by CrispEmbed run 40843704011; the file is copied unchanged from that run's retained artifact and checked against the receipt's size and SHA256.
2. Uploaded by CrispEmbed commit 017817a2eb13500aec2ec33d9ef86d69e3d7c1c7 (workflow ultrabert-publish.yml, protocol docs/ultrabert-publication-v1.json); manifest.json lists the pins, receipts and file hashes.

Usage

UltraBERT support is CPU-only, experimental and default-off:

CRISPEMBED_ULTRABERT=1 CRISPEMBED_ULTRABERT_QUANT=1 CRISPEMBED_ULTRABERT_UNICODE_PRETOKEN=1 CRISPEMBED_ULTRABERT_TRUNCATE=1 \
  crispembed -m ultrabert-embed-150m-q4_k-last-four-ffn.gguf "your text"