Modelpublic

anonymousresearch123/deka-qwen3-emb-0.6b-triplet-fact

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes20downloads

Model Card

deka-qwen3-emb-0.6b-triplet-fact

Full fine-tune of Qwen/Qwen3-Embedding-0.6B for Thai Supreme Court (deka) case retrieval.

—objective: triplet
—document register: fact (query side trained against fact text)
—best epoch: 1 (early-stopped on dev R@10 = 35.5)
—pooling: last token (left padding) + L2 normalize
—lr 2e-05, batch 16, tau 0.05, max_len doc 1300 / query 256, seed 42

Test results (full corpus, clean qrels)

eval register	R@5	R@10	R@50	R@100	ndcg@10	mrr
user	5.24	10.95	30.48	40.95	5.88	0.0741
lawer	14.76	23.33	46.67	62.38	14.99	0.1864
fact	22.38	34.76	59.52	73.33	21.88	0.2701

Usage

Queries must carry the instruction prefix used in training; documents must not.

python

from sentence_transformers import SentenceTransformer

m = SentenceTransformer("anonymousresearch123/deka-qwen3-emb-0.6b-triplet-fact")
q = m.encode(["ลูกจ้างถูกเลิกจ้างโดยไม่บอกกล่าวล่วงหน้า"], prompt_name="query")
d = m.encode(["<คำพิพากษาฎีกา ...>"])
print(m.similarity(q, d))

Plain transformers:

python

import torch, torch.nn.functional as F
from transformers import AutoModel, AutoTokenizer

tok = AutoTokenizer.from_pretrained("anonymousresearch123/deka-qwen3-emb-0.6b-triplet-fact", padding_side="left")
net = AutoModel.from_pretrained("anonymousresearch123/deka-qwen3-emb-0.6b-triplet-fact", torch_dtype=torch.bfloat16).eval()

INSTRUCT = "Given a legal case search query, retrieve relevant prior Supreme Court cases"
texts = [f"Instruct: {INSTRUCT}\nQuery: ลูกจ้างถูกเลิกจ้าง..."]   # query side only
enc = tok(texts, padding=True, truncation=True, max_length=1300, return_tensors="pt")
with torch.no_grad():
    h = net(**enc).last_hidden_state
emb = F.normalize(h[:, -1].float(), p=2, dim=-1)

Left padding is required — pooling takes the last position.

Training data

anonymousresearch123/deka-retrieval — splits train / eval / test. training_history.json in this repo holds the full per-epoch log and config.