TR Models Reverse Engineering: Trendyol-LLM + Cosmos-LLaMA + KanaryaTR
Turkey's open TR LLMs: Trendyol-LLM (Trendyol e-commerce-focused), Cosmos-LLaMA (Cosmos AI Lab), KanaryaTR (Boğaziçi NLP), TURNA, AnatoliaLLM. Reverse-engineering each: model card, training pipeline, base + data + technique. What can you take for yourself.
Şükrü Yusuf KAYA
30 min read
Advanced
1. TR Açık LLM Ailesi#
| Model | Geliştirici | Base | Lisans | TR-MMLU | Notlar |
|---|---|---|---|---|---|
| Trendyol-LLM 7B | Trendyol | Llama 2 7B | Apache 2.0 | 36.5 | E-ticaret focused, EN+TR mix |
| Cosmos-LLaMA-7B | Cosmos AI Lab | Llama 2 7B | MIT | 38.2 | TR continual pre-train + SFT |
| Cosmos-LLaMA-13B | Cosmos AI Lab | Llama 2 13B | MIT | 44.1 | aynı pipeline |
| KanaryaTR-Llama-3-8B | Boğaziçi NLP | Llama 3 8B | Apache 2.0 | 41.3 | academic, continual + SFT |
| TURNA-1B | Boğaziçi NLP | scratch | Apache 2.0 | 22.4 | TR-only pre-train (eski) |
| AnatoliaLLM-7B | Hacettepe | Mistral 7B | Apache 2.0 | 35.2 | academic |
| Aya Expanse 8B (TR'de iyi) | Cohere | scratch | CC-BY-NC | 42.3 | research only |
Cookbook'un kuralı (TR commercial): Cosmos-LLaMA-7B/13B veya KanaryaTR baseline; üzerine custom SFT.
2. Trendyol-LLM Reverse-Engineering#
Trendyol model card'tan okunabilen:
- Base: Llama 2 7B
- Pre-train extension: 25-50GB TR e-ticaret corpus (Trendyol kataloğu + reviews + customer service)
- SFT: 100K curated instruction (Trendyol-spesifik task'lar)
- DPO: 10K user feedback pair (canlı sistemden)
Cookbook'un derslik: Domain-specific TR LLM'leri base + domain-CPT + domain-SFT + DPO ile inşa et.
3. Cosmos-LLaMA Reverse-Engineering#
Cosmos AI Lab GitHub'da pipeline'ı açıkladı:
- Base: Llama 2 7B/13B
- Continual PT: 30GB TR corpus (Wiki + book + news + forum)
- SFT: 50K TR-Alpaca + curated 5K
- DPO: 10K AI feedback pair (Llama-3-70B judge)
- Eval: TR-MMLU, MT-Bench-TR, custom benchmark
Cookbook için kanıtladığı: Bu cookbook'un yapısı (Part IX + III + XI) Cosmos'un pipeline'ıyla çok yakın. Cosmos'tan farklı: cookbook RTX 4090 baseline, daha kompakt, modern modeller (Llama 3.x, Qwen3).
✅ Teslim
- Cosmos-LLaMA-7B model card'ını oku. 2) Kendi domain'in için Cosmos pipeline'ını adapt et. 3) Sonraki ders: 9.7 — TR Embedding & Reranker FT.
Yorumlar & Soru-Cevap
(0)Yorum yazmak için giriş yap.
Yorumlar yükleniyor...
Related Content
Part 0 — Engineering Foundations
Welcome to the Fine-Tuning Cookbook: System, Stage Taxonomy, and the Reproducibility Contract
Start LearningPart 0 — Engineering Foundations
Reproducibility Stack: Seeds, cuDNN Flags, and Deterministic CUDA — End the 'Works on My Machine' Problem
Start LearningPart 0 — Engineering Foundations