SATRI-4.0 Adapters & Merge Candidates (Qwen/Qwen3-4B · 36 Layer)

  • Pembuat: Pilus (PILUSATRIO / Satri)
  • Model Dasar: Qwen/Qwen3-4B (Revisi: 1cfa9a7208912126459214e8b04321603b3df60c, 36 Layer)
  • Model Rilis Utama (SATRI-4.0): PILUSATRIO/satri-4.0

Daftar Sub-Folder Adapter & Kandidat Merge (Generasi 4B)

  1. control_4b/ — Adapter kontrol multi-task 4B (rsLoRA $r=64, \alpha=128$ + NEFTune + T2e Batas Pengetahuan).
  2. spec_A_4b/ — Spesialis A 4B (Bahasa Indonesia, Budaya Nusantara, Instruksi L1–L2; Skor DEV Domain: 100,0%).
  3. spec_B_4b/ — Spesialis B 4B (Penalaran Matematika, Logika, & Sains + Self-Repair; Skor DEV Domain: 100,0%).
  4. spec_C_4b/ — Spesialis C 4B (Pemrograman, SQL, Algoritma, & Desain Web + Self-Repair; Skor DEV Domain: 100,0%).
  5. spec_D_4b/ — Spesialis D 4B (Function/Tool Calling <tool_call> & RAG + Self-Repair; Skor DEV Domain: 100,0%).
  6. satri_4b_uniform_ties/ — Pemenang Merge T5 (Uniform TIES 36-Layer, DEV 100,00%, TEST 93,75%, dipromosikan ke PILUSATRIO/satri-4.0).
  7. satri_4b_della_layerwise/ — Kandidat Merge Arm H (DELLA 36-Layer Stratified Merge [0..11], [12..27], [28..35], DEV 86,67% ketat / 100% semantik, TEST 87,50%, rata-rata panjang jawaban ringkas 161,0 token).
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for PILUSATRIO/satri-4b-adapters

Finetuned
Qwen/Qwen3-4B
Adapter
(1218)
this model