Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): (None, {}), NamedSplit('test'): ('json', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Sovereign Zenex MoE (DeepSeek-V3 inspired) LLM System (Local-Only)

This project implements a full local LLM stack aligned to your requirements:

  • No external AI APIs at runtime
  • No cloud AI services
  • Local-only training, fine-tuning, inference, evaluation, and monitoring
  • Zenex MoE (DeepSeek-V3 inspired) MoE/MLA/MTP design targets, config-driven

1) Hardware Targets

Full-scale target (reference)

  • 2,048x H100/H800 class GPUs
  • InfiniBand NDR + NVLink
  • 30–50 PB distributed storage

Practical bootstrap path

  • 8–16x H100/A100 for fine-tuning and pipeline validation
  • Scale out using the same codepaths and configs

2) Project Layout

3) Environment Setup

  1. Install Python 3.11+, Docker, NVIDIA driver/toolkit, and NCCL.
  2. Run:
    • bash scripts/setup.sh
  3. Activate venv:
    • source .venv/bin/activate
  4. Export package path:
    • export PYTHONPATH=$(pwd):$PYTHONPATH

4) Data Download (one-time)

Populate data/dataset_manifest.yaml with your approved sources. Then run:

  • python data/download_datasets.py --manifest data/dataset_manifest.yaml --output-root /mnt/nas/raw

5) Preprocessing + FIM + Repo Packing

  • python data/repo_packer.py --repo-root /mnt/nas/raw/repos --output /mnt/nas/clean/repo_docs.jsonl
  • python data/preprocess.py --input /mnt/nas/clean/repo_docs.jsonl --output /mnt/nas/clean/clean.jsonl
  • python data/fim_dataset.py --input /mnt/nas/clean/clean.jsonl --output /mnt/nas/clean/clean_fim.jsonl

6) Tokenizer

  • python tokenizer/train_tokenizer.py --data-root /mnt/nas/clean --manifest /mnt/nas/tokenizer/input_manifest.txt --output-prefix /mnt/nas/tokenizer/tokenizer --vocab-size 129280

7) Tokenization for Training

  • python data/tokenize_data.py --input /mnt/nas/clean/clean_fim.jsonl --tokenizer-model /mnt/nas/tokenizer/tokenizer.model --out-dir /mnt/nas/tokenized --general-ratio 0.87 --code-ratio 0.13

8) Pretraining

Edit training/config/training_config.yaml for cluster scale. Launch:

  • bash training/launch.sh

or directly:

  • python training/trainer.py --config training/config/training_config.yaml

9) SFT

  • python finetuning/sft_trainer.py --base-checkpoint /mnt/nas/checkpoints/base.pt --dataset /mnt/nas/sft/instructions.jsonl --tokenizer /mnt/nas/tokenizer/tokenizer.model --output /mnt/nas/checkpoints/sft.pt --epochs 1 --batch-size 1 --lr 1e-5

10) GRPO + Docker Execution Rewards

  • python finetuning/grpo_trainer.py --checkpoint /mnt/nas/checkpoints/sft.pt --tokenizer /mnt/nas/tokenizer/tokenizer.model --prompts /mnt/nas/rl/prompts.jsonl --output /mnt/nas/checkpoints/grpo.pt --group-size 8 --max-new-tokens 256

Sandbox smoke test:

  • python sandbox/run_sandbox_test.py

11) Inference (vLLM, OpenAI-compatible)

Edit inference/inference_config.yaml, then run:

  • python inference/serve.py --config inference/inference_config.yaml

Serves on local interface: http://0.0.0.0:8000

12) Evaluation

  • bash evaluation/evaluate_all.sh

Outputs:

  • evaluation/results/humaneval.json
  • evaluation/results/mbpp.json
  • evaluation/results/swebench.json
  • evaluation/results/summary.json

13) Monitoring (self-hosted)

Start stack:

  • bash monitoring/setup_monitoring.sh

Endpoints:

  • Prometheus: http://localhost:9090
  • Grafana: http://localhost:3000

Dashboards are in monitoring/grafana/dashboards.

14) Verification Checklist

  • No runtime external AI API calls
  • Local file-based model load/save paths
  • Local Docker sandbox execution with networking disabled
  • Local monitoring stack using Docker Compose
  • Local benchmark execution scripts

15) Notes on Scale

The architecture/config targets DeepSeek-V3 specs. Running at full 671B/37B-active scale requires multi-node H100 infrastructure. The same code paths are used from prototype scale to cluster scale via configuration.

Downloads last month
92