Jamjuri-EDGE-Preview-100-adapters

Cheva123/Jamjuri-EDGE-Preview-100-adapters

Jamjuri-EDGE-Preview-100-adapters is a machine learning model from Cheva123. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.

Downloads · 30 days

Access

Public

Updated Sep 23, 2026

Repo size

938 MB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors925 MB · 99%

At a glance

Library: peft
License: apache-2.0
Access: Public
Created: Sep 23, 2026
Updated: Sep 23, 2026
SHA: eb36c6b8

Base models

Cheva123/Jamjuri-EDGE-Preview-100

Model card

From the Hugging Face model README

Jamjuri Edge — Expert Adapters (E1 / E2 / E3 + Stage-1)

Raw LoRA adapters used to build the Jamjuri Edge family.

Series: part of JamjuriEDGE (4B) — collection · series card

✦ English

Overview

These are the raw LoRA adapters behind the merged models in Cheva123/Jamjuri-EDGE-Preview-100:

Folder	Adapter	Trained on	LoRA	Train loss	Size
`e1/`	E1 — Control / tools / IoT	`parents/jamjuri-edge-v4-stage1`	`r=16, α=32`	0.5326	126 MB
`e2/`	E2 — Math (Thai+EN)	same parent	`r=16, α=32`	0.5321	126 MB
`e3/`	E3 — Code	same parent	`r=16, α=32`	0.4332	126 MB
`stage1/`	Stage-1 — Primitive Core	Qwen3-4B-Instruct-2507	`r=64, α=128`	—	504 MB

All experts were trained from the same Stage-1 parent (non-thinking, assistant-only loss). Each folder also contains adapter_config.json, the PEFT README.md, and (for E1–E3) a train_report.json with the full training history.

Usage

from transformers import AutoModelForCausalLM
from peft import PeftModel

parent = AutoModelForCausalLM.from_pretrained(
    "Cheva123/Jamjuri-EDGE-Preview-100", subfolder="parents/jamjuri-edge-v4-stage1",
    dtype="bfloat16", device_map="auto")

model = PeftModel.from_pretrained(
    parent, "Cheva123/Jamjuri-EDGE-Preview-100-adapters", subfolder="e1"
).merge_and_unload()

Build your own merges

You can create your own: linear merges · TIES merges · DARE-TIES merges · custom expert ratios · custom global lambda. Reference implementation (TIES, adapter-aware): recipes/stage2_ties.py in the main repo.

Main repo: Cheva123/Jamjuri-EDGE-Preview-100
Datasets: Cheva123/jamjuri-edge-v4-stage2-datasets

✦ ภาษาไทย

ภาพรวม

นี่คือ LoRA adapter ตัวดิบ ที่อยู่เบื้องหลังโมเดล merged ใน Cheva123/Jamjuri-EDGE-Preview-100

โฟลเดอร์	Adapter	เทรนบน	LoRA	Train loss	ขนาด
`e1/`	E1 — Control / tools / IoT	`parents/jamjuri-edge-v4-stage1`	`r=16, α=32`	0.5326	126 MB
`e2/`	E2 — Math (ไทย+อังกฤษ)	parent เดียวกัน	`r=16, α=32`	0.5321	126 MB
`e3/`	E3 — Code	parent เดียวกัน	`r=16, α=32`	0.4332	126 MB
`stage1/`	Stage-1 — Primitive Core	Qwen3-4B-Instruct-2507	`r=64, α=128`	—	504 MB

expert ทั้ง 3 ตัวเทรนจาก Stage-1 parent เดียวกัน (โหมด non-thinking, คิด loss เฉพาะคำตอบ assistant) · ในแต่ละโฟลเดอร์มี adapter_config.json, PEFT README.md และ (E1–E3) train_report.json ที่เก็บประวัติการเทรนครบ

วิธีใช้งาน

from transformers import AutoModelForCausalLM
from peft import PeftModel

parent = AutoModelForCausalLM.from_pretrained(
    "Cheva123/Jamjuri-EDGE-Preview-100", subfolder="parents/jamjuri-edge-v4-stage1",
    dtype="bfloat16", device_map="auto")

model = PeftModel.from_pretrained(
    parent, "Cheva123/Jamjuri-EDGE-Preview-100-adapters", subfolder="e1"
).merge_and_unload()

เอาไป merge เอง

ทำได้ตามใจ ทั้งแบบ linear, TIES, DARE-TIES หรือปรับ ratio / lambda เอง — โค้ดอ้างอิง (TIES แบบอ่าน adapter ตรงๆ): recipes/stage2_ties.py ใน main repo

Main repo: Cheva123/Jamjuri-EDGE-Preview-100
Datasets: Cheva123/jamjuri-edge-v4-stage2-datasets

Library: peft
License: apache-2.0
Languages: th, en
Created: Sep 23, 2026
Updated: Sep 23, 2026