Анализ Π΄Π°Π½Π½Ρ‹Ρ… (Data analysis)

@data_analysis_ml Β· ai Β· Russian Β· active

Subscribers 52K Subscribers, last 30 days: 52K β†’ 52K +0 24h
Reach per post (ERR) 11.1%5.8K avg views
Posting 4.9/day last 2 hours ago
Reactions per 1K views 31.3 ❀ πŸ‘ πŸ”₯

πŸ”₯ Soup - Π΄ΠΎΠΎΠ±ΡƒΡ‡Π΅Π½ΠΈΠ΅ LLM ΠΎΠ΄Π½ΠΎΠΉ ΠΊΠΎΠΌΠ°Π½Π΄ΠΎΠΉ ΠŸΡ€ΠΎΠ΅ΠΊΡ‚ ΡƒΠΏΡ€ΠΎΡ‰Π°Π΅Ρ‚ fine-tuning ΠΈ post-training ΠΌΠΎΠ΄Π΅Π»Π΅ΠΉ Π΄ΠΎ схСмы: ΠΎΠ΄ΠΈΠ½ YAML-ΠΊΠΎΠ½Ρ„ΠΈΠ³ β†’ ΠΎΠ΄Π½Π° ΠΊΠΎΠΌΠ°Π½Π΄Π°. Π‘Π°ΠΌΠΎΠ΅ интСрСсноС: - ΠΌΠΎΠΆΠ½ΠΎ Π΄ΠΎΠΎΠ±ΡƒΡ‡Π°Ρ‚ΡŒ 8B-модСль Π΄Π°ΠΆΠ΅ Π½Π° Π½ΠΎΡƒΡ‚Π±ΡƒΠΊΠ΅ с 4 Π“Π‘ VRAM - layer streaming ΠΏΠΎΠ΄Π°Ρ‘Ρ‚ слои ΠΌΠΎΠ΄Π΅Π»ΠΈ Π½Π° GPU ΠΏΠΎ ΠΎΠ΄Π½ΠΎΠΌΡƒ - Ρ€Π°Π±ΠΎΡ‚Π°Π΅Ρ‚ с QLoRA ΠΈ Π»ΠΎΠΊΠ°Π»ΡŒΠ½Ρ‹ΠΌΠΈ GPU - автоматичСски ΠΏΠΎΠ΄Π±ΠΈΡ€Π°Π΅Ρ‚ batch size, устройство ΠΈ ΠΊΠ²Π°Π½Ρ‚ΠΎΠ²Π°Π½ΠΈΠ΅ - Π΅ΡΡ‚ΡŒ DPO, ORPO, SimPO ΠΈ KTO - Π±Π΅Π· SSH ΠΈ ΠΎΡ‚Π΄Π΅Π»ΡŒΠ½ΠΎΠΉ ΠΎΠ±Π»Π°Ρ‡Π½ΠΎΠΉ инфраструктуры На RTX 3050 Laptop 4 GB Π°Π²Ρ‚ΠΎΡ€Ρ‹ запускали Llama 3.1 8B Instruct с ΠΏΠΈΠΊΠΎΠ²Ρ‹ΠΌ ΠΏΠΎΡ‚Ρ€Π΅Π±Π»Π΅Π½ΠΈΠ΅ΠΌ ΠΎΠΊΠΎΠ»ΠΎ 3,32 Π“Π‘ VRAM. https://github.com/MakazhanAlpamys/Soup

Views over time

02K4K6K28 Aug 10:1012 Nov 06:465.1K28 Aug 10:10: 0 views12 Nov 06:46: 5,120 views
Snapshots we took after posting.
WhenViews
28 Aug 2026, 10:100
12 Nov 2026, 06:465.1K

Who shared this post

Posts in other channels we index that link to this one. (Forwards with a hidden source can't be traced.)

Nobody yet.