docs: Best datasets guide with verified formats + proven recipes

ai-ml/hf-finetuning/BEST_DATASETS.md

@@ -0,0 +1,72 @@
+# Best Datasets for SFT Fine-Tuning — Verified Guide
+
+## Dataset Rankings (Quality → Model Performance)
+
+### #1: allenai/tulu-3-sft-mixture — THE BEST
+- **Size**: 939K examples from 19 curated sources
+- **Format**: messages column (role/content) - ZERO PREPROCESSING
+- **Sources**: FLAN v2, Persona MATH, Evol CodeAlpaca, WildChat, Aya, NuminaMath, WildGuard, WildJailbreak, no_robots, OASST1, SciRIFF, etc.
+- **Proven Results on Llama-3.1-8B**: MMLU 53.5, GSM8K 79.9, IFEval 63.6, HumanEval 76.8
+- **Training Recipe**: LR=5e-6, batch=128, epochs=2, max_seq=4096, linear schedule
+- **Status**: VALIDATED - column format confirmed via hf_inspect_dataset
+
+### #2: open-thoughts/OpenThoughts-114k — REASONING CoT
+- **Size**: 114K examples with DeepSeek-R1 reasoning traces
+- **Format**: conversations column (from/value ShareGPT) - NEEDS CONVERSION
+- **Best For**: Math, code, science with chain-of-thought
+- **Conversion**: See train_openthoughts.py
+- **Training Recipe**: LR=2e-4, batch=16, epochs=2, cosine schedule
+- **Status**: VALIDATED - format confirmed, converter tested
+
+### #3: HuggingFaceH4/ultrachat_200k — GENERAL CHAT
+- **Size**: 208K multi-turn conversations
+- **Format**: messages column - ZERO PREPROCESSING (use train_sft split)
+- **Best For**: General conversational ability
+- **Training Recipe**: LR=2e-4, batch=16, epochs=1
+
+### #4: mlabonne/FineTome-100k — CURATED COMPACT
+- **Size**: 100K quality-scored examples
+- **Format**: conversations (ShareGPT) - NEEDS CONVERSION
+- **Best For**: Quick fine-tune with curated quality
+
+### #5: HuggingFaceH4/no_robots — HUMAN-WRITTEN
+- **Size**: 9.5K examples (all human-written)
+- **Format**: messages column - ZERO PREPROCESSING
+- **Best For**: High-quality instruction following
+
+## How to Train
+
+### Full Training (Tulu 3 - 940K) — A100 80GB, ~6h
+```
+python ai-ml/hf-finetuning/train_tulu3.py
+```
+
+### Reasoning Training (OpenThoughts - 114K) — A100 80GB, ~2h
+```
+python ai-ml/hf-finetuning/train_openthoughts.py
+```
+
+### Quick Test (100 steps) — Any GPU
+```
+python ai-ml/hf-finetuning/train_tulu3.py --max_steps 100 --no_push
+```
+
+## LoRA Config (LoRA Without Regret - Schulman 2025)
+
+| Parameter | Tulu 3 Recipe | OpenThoughts Recipe |
+|-----------|---------------|---------------------|
+| lora_r | 256 | 256 |
+| lora_alpha | 16 | 16 |
+| target_modules | all-linear | all-linear |
+| learning_rate | 5e-6 | 2e-4 |
+| effective_batch | 128 | 16 |
+| epochs | 2 | 2 |
+| max_seq_length | 4096 | 4096 |
+| lr_schedule | linear | cosine |
+| packing | True (bfd_split) | True (bfd_split) |
+| assistant_only_loss | True | True |
+
+## Key Research Sources
+- Tulu 3: allenai/Llama-3.1-Tulu-3-8B-SFT model card
+- LoRA Without Regret: Schulman et al., 2025
+- Data quality > quantity: arXiv 2402.05123