SkillRet — Agent Skill Retrieval - a ThakiCloud Collection

ThakiCloud 's Collections

Thaki Energy — Short-Term Electricity Load Forecasting

Thaki TrustEdge — Tiny Korean Command Guards

Thaki-Fraud — Synthetic Transaction Anomaly Baselines

CareQ — Ultra-Low-Bit LLM Quantization

Thaki Sovereign AI — Korean Domain Specialist Models

Thaki Sec — Security Models

Task-Specific Rerankers

ThakiCloud Bolt E4B Specialists

Thaki Speech — Korean ASR & TTS

Thaki-Ops — Kubernetes SRE Agent Models

KoToolAct — Korean Enterprise Tool Calling 4B

Thaki-Decision — Generation-free LLM routing

Ko-Voice-Command — Korean Spoken Command Understanding

FactGuard — Evidence-Grounded Verification

Multilingual Embedding — Low-Bit Quantization

Commerce AI Models

Thaki-PolicyGate — Korean regulation-condition checking

ThakiCloud Security Models

Thaki-HW — Hardware Design LLM (Research Preview)

EmbeddingGemma 2 Edge — Quantization

RAG-Gate — Answer / Retrieve / Stop Gates for RAG

ChainCheck — Chain-Sensitive Evidence Judges

PACE — Enterprise Policy Decisions: Learned or Reasoned

Merchant Desk — Korean Small-Business Customer Desk

Qwen-Image-2.1 — Few-Step Distillation

K-Decision — Korean typed-decision adapters (weights only)

K-MT — Korean-Specialized Translation Without Forgetting

Satoori-KO — Korean Regional Dialect 27B

Evals & Reproducibility

Language Confusion Suppression — Vocabulary Pruning

Vision-Language — Serving Quantization

Qwen3-30B-A3B — Serving Quantization

Qwen3-Coder-30B-A3B — Serving Quantization

Qwen3.8-27B — Serving Quantization

SkillRet — Agent Skill Retrieval

Human-KO — Korean Style-Aligned 27B

SkillRet — Agent Skill Retrieval

updated 13 days ago

Retrieval embeddings, benchmark, and index that let an LLM agent find the right skill inside a library of 1,000+ skills.

Upvote

ThakiCloud/SKILLRET-Embedding-8B

Sentence Similarity • 8B • Updated Sep 25 • 145 • 4

Note High-precision retrieval (8B)
ThakiCloud/SKILLRET-Embedding-0.6B

Sentence Similarity • 0.6B • Updated Sep 25 • 780 • 11

Note On-device, low-latency retrieval (0.6B)
ThakiCloud/superskillret-onnx-int8

Feature Extraction • Updated Sep 10 • 9

Note ONNX INT8 — for CPU inference
ThakiCloud/SKILLRET

Viewer • Updated Sep 25 • 218k • 4.93k • 9

Note Training and evaluation benchmark
ThakiCloud/superskillret-index

Viewer • Updated Sep 7 • 16.8k • 177

Note Prebuilt skill index
SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents

Paper • 2605.05726 • Published May 7 • 1

Note Technical report
ThakiCloud/SKILLRET-Edge-22M

Feature Extraction • 22.7M • Updated Sep 16 • 291

Note Distilled 22M CPU student (fp16)
ThakiCloud/SKILLRET-Edge-22M-int4

Feature Extraction • 22.7M • Updated Sep 16 • 101

Note 22M student, uniform int4 g16 (17 MB)
ThakiCloud/SKILLRET-Edge-109M

Feature Extraction • 0.1B • Updated Sep 16 • 118

Note Distilled 109M CPU student (fp16)
ThakiCloud/SKILLRET-Edge-109M-int3

Feature Extraction • 0.1B • Updated Sep 16 • 99

Note 109M student, uniform int3
ThakiCloud/SKILLRET-Edge-22M-int3

Feature Extraction • 22.7M • Updated Sep 16 • 16

Note 22M student, uniform int3 g16 (14 MB)
ThakiCloud/SKILLRET-Edge-22M-drift-mixed

Feature Extraction • 22.7M • Updated 26 days ago • 119

Note 22M student, drift-allocated mixed precision (14 MB, int3 bytes)
ThakiCloud/SKILLRET-Edge-109M-int4

Feature Extraction • 0.1B • Updated Sep 16 • 91

Note 109M student, uniform int4
ThakiCloud/SKILLRET-Edge-109M-int3-g32

Feature Extraction • 0.1B • Updated Sep 16 • 16

Note 109M student, uniform int3 g32
Where Post-Training Quantization Breaks Text Embedders: A Measured Map Across Four Embedder Families

Paper • 2609.16391 • Published Sep 14

Note Uniform-PTQ map behind the Edge line. SKILLRET-Embedding-0.6B and the distilled 109M student are two of its five checkpoints across four embedder families. Protecting the embedding table and allocating bits by module sensitivity do not transfer as stated.
Quantize by Drift: Label-Free Mixed-Precision Post-Training Quantization for Text Embedders

Paper • 2610.09227 • Published 28 days ago

Note Label-free drift criterion behind SKILLRET-Edge-22M-drift-mixed. Orders mixed-precision plans against held-out retrieval quality at macro Spearman 0.911 with no labels, but on three sealed held-out embedders it lands below same-budget uniform at the main budget (-0.99, -0.85, -1.01 points).
ThakiCloud/SKILLRET-Edge-Arctic-305M

Sentence Similarity • 0.3B • Updated 20 days ago • 20

Note Distilled 305M multilingual CPU student, FP32 ONNX (en/ko/mixed)
When PTQ Calibration Helps or Hurts: Dose-Driven Collapse and Validation-Gated Abstention in Quantized Text Embedders

Paper • 2610.27397 • Published 15 days ago

Note Third paper of the embedder-quantization line: learned PTQ calibration on e5-base-v2, BGE-M3 and EmbeddingGemma-300M. Outcome tracks update dose, from +1.47 to -71.53 NDCG@10 on one grid; a validation-gated accept rule made 0 harmful picks in 30 decisions but paid 1.52 pp mean regret. The SKILLRET-Edge quantized checkpoints use no learned calibration.

Upvote