What ships here

• train_stage1.py for the Stage 1 GPT baseline
• train_llama_stage2.py for LoRA fine-tuning on LLaMA
• train_lora.py as a compatibility entry point for the LoRA path
• train_pipeline.py to run both stages in sequence
• build_datasets.py to generate topic-sharded synthetic datasets
• data/ with example JSONL training sets

Model Direction

The project is intended to produce two artifacts:

• a meaningful internal baseline from Stage 1
• a higher-quality assistant checkpoint or adapter from Stage 2

The Stage 2 model is the release artifact for normal inference.

Base Model

Stage 2 currently targets:

• meta-llama/Llama-3.1-8B-Instruct

The earlier LoRA script in this repo also supports the same model family.

Training Flow

Recommended flow:

1. Run python build_datasets.py --output_dir data/generated --include_docs to generate topic-specific shards and aggregate JSONL files.
2. Train Stage 1 with data/generated/stage1_sft.jsonl if you want the GPT baseline.
3. Run python train_stage1.py to build the small baseline.
4. Train Stage 2 with data/generated/stage2_conrad_sft.jsonl or let python train_pipeline.py --include_docs do the build step automatically.
5. Merge the Stage 2 adapter with python merge_stage2_lora.py.
6. Sync the merged checkpoint into the model repo with python sync_checkpoint.py.
7. Publish the merged checkpoint.

Intended Use

This project is designed for:

• conversational assistants
• documentation assistants
• support routing
• enterprise workflows
• knowledge assistants
• internal tooling
• structured response generation

Notes

• Stage 1 and Stage 2 are separate training jobs.
• Stage 1 is not a wrapper around LLaMA.
• Stage 2 is the higher-quality assistant tuning step.
• The repo includes example datasets only; real training data is still required.
• Set CONRAD_ENDPOINT_URL and HF_TOKEN in the Space secrets or environment to enable production chat.
• If the endpoint is unavailable, the Space will show a clear fallback instead of generating from the raw checkpoint.