cudo528's picture
Update README.md
3ced6dc verified
|
Raw History Blame Contribute Delete
3.74 kB
metadata
license: apache-2.0
tags:
  - merge
  - evolutionary-merge
  - darwin
  - darwin-v5
  - model-mri
  - reasoning
  - advanced-reasoning
  - chain-of-thought
  - thinking
  - qwen3.5
  - qwen
  - claude-opus
  - distillation
  - multilingual
  - benchmark
  - open-source
  - apache-2.0
  - layer-wise-merge
  - coding-agent
  - tool-calling
  - long-context
  - mlx
  - quantized
language:
  - en
  - zh
  - ko
  - ja
  - de
  - fr
  - es
  - ru
  - ar
  - multilingual
pipeline_tag: text-generation
library_name: mlx
base_model: FINAL-Bench/Darwin-9B-Opus
quantized_by: cudo528
bits: 4
model-index:
  - name: Darwin-9B-Opus
    results:
      - task:
          type: question-answering
          name: Question Answering
        dataset:
          name: GPQA Diamond
          type: Idavidrein/gpqa
          config: gpqa_diamond
        metrics:
          - type: accuracy
            value: 82.5
            name: Accuracy

Darwin-9B-Opus (MLX 4-bit Optimized)

This repository provides an MLX-optimized 4-bit quantized version of FINAL-Bench/Darwin-9B-Opus, an ultra-efficient advanced reasoning model.

์ง„ํ™”์  ๋จธ์ง€(Evolutionary Merge) ๋ฐ ๋ ˆ์ด์–ด๋ณ„ ๋จธ์ง€(Layer-wise Merge) ๊ธฐ๋ฒ•์ด ์ ์šฉ๋œ 9B ์ฒด๊ธ‰์˜ ๊ฐ•๋ ฅํ•œ ๋ชจ๋ธ์„ Apple Silicon(Mac) ํ™˜๊ฒฝ์—์„œ ํ•˜๋“œ์›จ์–ด ๋ฆฌ์†Œ์Šค๋ฅผ ์ตœ์†Œํ™”ํ•˜๋ฉฐ ๊ตฌ๋™ํ•  ์ˆ˜ ์žˆ๋„๋ก 4-bit๋กœ ์–‘์žํ™” ๋ฐ ํฌ๋งท ๋ณ€ํ™˜์„ ์™„๋ฃŒํ•œ ๋ฒ„์ „์ž…๋‹ˆ๋‹ค.

โš ๏ธ Hugging Face View Note: ํ—ˆ๊น…ํŽ˜์ด์Šค ์‹œ์Šคํ…œ์ด MLX 4๋น„ํŠธ ํŒจํ‚น ํฌ๋งท(U32/BF16)์„ ํŒŒ์‹ฑํ•˜๋Š” ๊ณผ์ •์—์„œ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜ ๋ฐ ํ…์„œ ํƒ€์ž…์ด ์ผ์‹œ์ ์œผ๋กœ ๋‹ค๋ฅด๊ฒŒ ํ‘œ๊ธฐ๋  ์ˆ˜ ์žˆ์œผ๋‚˜, ์‹ค์ œ 4๋น„ํŠธ ์–‘์žํ™” ๊ฐ€์ค‘์น˜๊ฐ€ ์ •์ƒ ์ ์šฉ๋œ ํŒŒ์ผ์ž…๋‹ˆ๋‹ค. ๋งฅ๋ถ ๋กœ์ปฌ ํ™˜๊ฒฝ์—์„œ ์ •์ƒ์ ์ธ ๋ฉ”๋ชจ๋ฆฌ ์ ์œ ์œจ๋กœ ๊ตฌ๋™๋ฉ๋‹ˆ๋‹ค.

๐Ÿš€ Key Features

  • Extreme Efficiency (4-bit): 9B ์ฒด๊ธ‰์˜ ๊ฐ€์ค‘์น˜๋ฅผ 4๋น„ํŠธ๋กœ ์ •๊ตํ•˜๊ฒŒ ์••์ถ•ํ•˜์—ฌ, VRAM(ํ†ตํ•ฉ ๋ฉ”๋ชจ๋ฆฌ)์ด ์ ์€ Mac ๊ธฐ๊ธฐ์—์„œ๋„ ๊ฐ€๋ณ๊ณ  ๋น ๋ฅด๊ฒŒ ์ถ”๋ก (Inference)์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
  • Top-Tier 9B Reasoning: ํ•˜๋“œ์ฝ”์–ด ๋ฐ•์‚ฌ๊ธ‰ ์ถ”๋ก  ํ…Œ์ŠคํŠธ์ธ **GPQA Diamond์—์„œ 82.5%**๋ผ๋Š” ์ฒด๊ธ‰ ํŒŒ๊ดด ์ˆ˜์ค€์˜ ๊ณ ๋“์ ์„ ๊ธฐ๋กํ–ˆ์Šต๋‹ˆ๋‹ค. ๊นŠ์€ ์‚ฌ๊ณ  ๊ณผ์ •(<thought>)์„ ํ†ตํ•œ ๊ณ ๋‚œ๋„ ๋…ผ๋ฆฌ ์ถ”๋ก ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.
  • Agentic Capabilities: ์ฝ”๋”ฉ ์—์ด์ „ํŠธ(Coding Agent) ์„ฑ๋Šฅ๊ณผ ์ •๊ตํ•œ ๋„๊ตฌ ํ˜ธ์ถœ(Tool Calling) ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ๊ธฐ๋ณธ ๋‚ด์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
  • Global Multilingual: ํ•œ๊ตญ์–ด, ์˜์–ด, ์ค‘๊ตญ์–ด, ์ผ๋ณธ์–ด, ํ”„๋ž‘์Šค์–ด, ๋…์ผ์–ด ๋“ฑ 10๊ฐœ ์ด์ƒ์˜ ๋‹ค๊ตญ์–ด ํ™˜๊ฒฝ์„ ์™„๋ฒฝํ•˜๊ฒŒ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ› ๏ธ Installation

์ด ๋ชจ๋ธ์„ ๋กœ์ปฌ ํ™˜๊ฒฝ์—์„œ ๊ตฌ๋™ํ•˜๋ ค๋ฉด ์ตœ์‹  ๋ฒ„์ „์˜ mlx-lm ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๊ฐ€ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.

pip install -U mlx-lm

๐Ÿ’ป How to Use

1. Python API๋ฅผ ์ด์šฉํ•œ ์ถ”๋ก  ์ฝ”๋“œ

from mlx_lm import load, generate

# ๋ณธ์ธ์˜ ํ—ˆ๊น…ํŽ˜์ด์Šค ์ €์žฅ์†Œ ์•„์ด๋””/์ด๋ฆ„์œผ๋กœ ์ˆ˜์ •ํ•˜์—ฌ ์‚ฌ์šฉํ•˜์„ธ์š”.
model, tokenizer = load("cudo528/Darwin-9B-Opus-mlx-4bit")

prompt = "์ˆ˜ํ•™์  ๊ท€๋‚ฉ๋ฒ•์˜ ์›๋ฆฌ๋ฅผ ์ค‘ํ•™์ƒ๋„ ์ดํ•ดํ•  ์ˆ˜ ์žˆ๊ฒŒ ์‰ฌ์šด ๋น„์œ ๋ฅผ ๋“ค์–ด ํ•œ๊ตญ์–ด๋กœ ์„ค๋ช…ํ•ด์ค˜."

response = generate(
    model, 
    tokenizer, 
    prompt=prompt, 
    max_tokens=2048,
    verbose=True # ๋ชจ๋ธ์˜ ์‹ฌ์ธต ์‚ฌ๊ณ  ๊ณผ์ •(<thought>)์„ ํ„ฐ๋ฏธ๋„์—์„œ ์‹ค์‹œ๊ฐ„์œผ๋กœ ๋ณด๋ ค๋ฉด True
)
print(response)

2. ํ„ฐ๋ฏธ๋„(CLI) ํ™˜๊ฒฝ์—์„œ ์ฆ‰์‹œ ์‹คํ–‰

mlx_lm.generate --model cudo528/Darwin-9B-Opus-mlx-4bit --prompt "Write a Python script for a simple custom neural network layer." --max-tokens 1024

๐Ÿ“‚ Architecture & Lineage

  • Original Base Model: FINAL-Bench/Darwin-9B-Opus
  • Quantization Bit: 4-bit
  • Format: MLX Formatted (.safetensors packed)
  • License: Apache-2.0