Add research rationale document

RESEARCH.md

@@ -0,0 +1,69 @@
+# Nomen-AI Research Rationale
+
+Nomen-AI treats name generation as controllable cross-lingual morpho-phonetic synthesis rather than unconstrained next-token completion.
+
+## Why not a 7B+ LLM?
+
+The target runtime is a free-tier Google Colab T4 with about 15GB VRAM. A 7B+ model is not appropriate for reliable full training. Nomen-AI therefore uses:
+
+- `Qwen/Qwen2.5-1.5B-Instruct`
+- LoRA adapters
+- fp16
+- short sequence length (`max_length=192`)
+
+## Control-token conditioning
+
+Control tokens are based on the CTRL-style idea that prepended codes condition generation:
+
+- Paper: CTRL: A Conditional Transformer Language Model for Controllable Generation
+- arXiv: https://arxiv.org/abs/1909.05858
+
+Nomen-AI controls:
+
+```text
+[ROOT:japanese:40+nordic:60] [THEME:gaming] [SYL:3] [LEN:8] [CREATIVE:0.8]
+```
+
+## SFT phase
+
+TRL SFTTrainer is used to teach prompt-to-name behavior from the synthetic morpheme dataset.
+
+Docs: https://huggingface.co/docs/trl/sft_trainer
+
+Dataset: https://huggingface.co/datasets/krystv/nomen-ai-sft
+
+## DPO phase
+
+DPO steers away from generic/derivative names using preference pairs:
+
+- chosen: novel morpho-phonetic name
+- rejected: derivative/cliche name (`TechHub`, `Brandify`, `GetZone`, etc.)
+
+Paper: https://arxiv.org/abs/2305.18290
+Docs: https://huggingface.co/docs/trl/dpo_trainer
+Dataset: https://huggingface.co/datasets/krystv/nomen-ai-dpo
+
+## Creativity knob
+
+Low creativity uses contrastive search:
+
+- SimCTG: https://arxiv.org/abs/2202.06417
+- Contrastive Search Is What You Need: https://arxiv.org/abs/2210.14140
+
+High creativity uses min-p sampling:
+
+- Min-p sampling: https://arxiv.org/abs/2407.01082
+
+## Anti-duplication
+
+Nomen-AI includes an inference-time novelty matrix:
+
+- fuzzy string similarity
+- character n-gram overlap
+- thresholded novelty score
+
+Implementation: https://huggingface.co/krystv/nomen-ai/blob/main/nomen_ai/antidup.py
+
+## Practical status
+
+The code, datasets, notebooks, demo, Docker path, and validation scripts are complete. Actual SFT/DPO adapter weights are not present yet because GPU/HF Jobs execution from the agent environment was repeatedly rejected.