# Nomen-AI Final Report

## Mission

Build an end-to-end, production-ready, T4-compatible pipeline for controllable cross-lingual morpho-phonetic brand/channel name synthesis.

## Completed public assets

| Asset | URL |
|---|---|
| Main repo | https://huggingface.co/krystv/nomen-ai |
| SFT dataset | https://huggingface.co/datasets/krystv/nomen-ai-sft |
| DPO dataset | https://huggingface.co/datasets/krystv/nomen-ai-dpo |
| SFT adapter target | https://huggingface.co/krystv/nomen-ai-sft-lora |
| DPO adapter target | https://huggingface.co/krystv/nomen-ai-dpo-lora |
| Demo Space | https://huggingface.co/spaces/krystv/nomen-ai-demo |

## Architecture

- Base model: `Qwen/Qwen2.5-1.5B-Instruct`
- Fine-tuning: LoRA + TRL SFTTrainer
- Preference tuning: TRL DPOTrainer
- Controls: `ROOT`, `THEME`, `SYL`, `LEN`, `CREATIVE`
- Anti-duplication: fuzzy similarity + character n-gram overlap
- Creativity decoding: contrastive search for low creativity, min-p sampling for high creativity

## Completed engineering

- Modular Python package under `nomen_ai/`
- Synthetic dataset builder
- SFT training script
- DPO training script
- Smoke test script
- Evaluation script
- Artifact checker
- Adapter card updater
- CPU validation tests
- Colab notebooks
- Dockerfile and docker-compose GPU training path
- Makefile command map
- Gradio demo Space
- Research/citation/license docs

## Datasets

### SFT

- Repo: https://huggingface.co/datasets/krystv/nomen-ai-sft
- Format: TRL `messages`
- Rows: 12,000 generated rows

### DPO

- Repo: https://huggingface.co/datasets/krystv/nomen-ai-dpo
- Format: TRL `prompt`, `chosen`, `rejected`
- Rows: 6,000 preference rows

## Execution status

The following were attempted but could not be completed from the agent environment because GPU/HF Jobs execution was repeatedly rejected:

1. GPU sandbox creation
2. HF Jobs T4 smoke test
3. Retried HF Jobs smoke test
4. Additional HF Jobs validation/training attempts

Current artifact state:

- `krystv/nomen-ai-sft-lora`: repo exists, no adapter weights yet
- `krystv/nomen-ai-dpo-lora`: repo exists, no adapter weights yet

## How to complete training

### Colab T4

```bash
git clone https://huggingface.co/krystv/nomen-ai
cd nomen-ai
pip install -q -r requirements.txt
huggingface-cli login
bash scripts/train_all_colab.sh
```

### Docker GPU

```bash
git clone https://huggingface.co/krystv/nomen-ai
cd nomen-ai
export HF_TOKEN=hf_...
docker compose up --build
```

## Expected trained outputs

- `krystv/nomen-ai-sft-lora/adapter_model.safetensors`
- `krystv/nomen-ai-dpo-lora/adapter_model.safetensors`

## Live demo

The current demo is CPU-safe and uses the morpheme synthesizer fallback:

https://huggingface.co/spaces/krystv/nomen-ai-demo

It displays live artifact status and can be upgraded to model-backed inference after DPO weights are present.