Aligned 2k-step further-training from ckpt-2000 (MCQA dropped; strong+amendment upsampled)

f2eaf53 verified 9 days ago

1.08 kB

base_model: Qwen/Qwen3-Embedding-0.6B
library_name: transformers
pipeline_tag: feature-extraction
tags:
  - qwen3
  - embeddings
  - legal-retrieval
  - procurement
  - lora-merged

LorMolf/Qwen-Embedding-ProcCode-aligned-2k

Merged Qwen3-Embedding checkpoint for Italian public-procurement retrieval.

Base model: Qwen/Qwen3-Embedding-0.6B
Merge base model: src_appalti/src_retriever/data/qwen3_embedding/merged/Qwen-Embedding-ProcCode-checkpoint-2000
Initialization adapter checkpoint: src_appalti/src_retriever/data/qwen3_embedding/merged/Qwen-Embedding-ProcCode-checkpoint-2000
Adapter checkpoint: src_appalti/src_retriever/data/qwen3_embedding/outputs/qwen3-embedding-0_6b-basecode-aligned-2k-20260623_232914/v0-20260623-233753/checkpoint-2000
Merge time: 2026-06-24T08:48:00.004976+00:00
Training backend: SWIFT qwen3_emb LoRA, InfoNCE
Expected query format: Instruct: <retrieval instruction>\nQuery: <question>
Document format: raw article/source or wiki-node text without instruction prefix
Max context used during training/eval: 32768 tokens