LateOn-Code-edge

The LateOn-Code collection is composed of PyLate models optimized for code retrieval. These late interaction models are first pre-trained following the methodology of CoRNStack. These pre-trained models are then further fine-tuned on train sets of CoIR using the nv-retriever methodology to mine hard negatives while preventing false negatives.

We started from the two best ColBERT models on the BEIR benchmark for their respective sizes. The first one, LateOn-Code is based on in-house LateOn model, a new version of GTE-ModernColBERT-v1 built on ModernBERT-base (also developed at LightOn). This version underwent significantly deeper training, crossing the 57 mark on BEIR, almost a 2.5-point improvement and is thus SOTA by a large margin. We'll release this base model along with training data and boilerplates in the near future, so stay tuned! The second, LateOn-Code-edge is a smaller model based on the edge-colbert model family from mixedbread, using the smallest variant (Ettin-17M) for maximum efficiency. For more details on the training setup, please refer to our blogpost.

The original CoRNStack data in a format compatible with PyLate can be found here while the fine-tuning data can be found here. Training boilerplates can be found here in the PyLate repository

MTEB (Code, v1) benchmark results

Pre-trained models achieve very competitive results as the 17M model outperforms the very strong granite-embedding-small-english-r2 by an average of 1.7. This is truly impressive, as the granite model is almost three times bigger (17M vs 48M), but is also a beast on its own in the <100M parameters range. It also outperforms the larger granite variant (149M). The larger version nicely scales by improving over the performance of its little sibling by 6.5 on average.

Although the pre-training results are already very impressive given that they are mostly out-of-domain, running a proper fine-tuning using the training data of CoIR significantly boost the performance of the models. Notably, the 17M model increases from 57.50 to 66.64 (+9.14), getting pretty close to EmbeddingGemma-300M while being 17 times smaller. The larger one increases from 63.77 to 74.12 (+10.35), strongly outperforming EmbeddingGemma-300M and getting closer to strong LLM models such as Qwen3-Embedding-0.6B and C2LLM-0.5B while being much smaller.