Skip to content

Model artifact list and front-end learning parameter mapping

[!NOTE] For the latest implementation status, please refer to Functional Implementation Status (Remaining Functionality).

Implementation notes (artifacts): See docs/implementation/ARTIFACT_MANIFESTS.md for the artifact_manifest.json output by the training script and recommended CLI flags.

Creation date: 2025-12-21

This document determines and lists "implementable model artifact names" for the 24-node configuration defined previously. Also, clarify which artifact settings correspond to the main parameters specified in the front-end learning form (LLM/encoder learning).

1. Model artifact candidates by node

  • Observation node (Sensing x4)
  • sensing-camera-preproc-v1 (image preprocessing pipeline)
  • sensing-audio-preproc-v1 (audio preprocessing pipeline)

  • sensing-iot-normalizer-v1 (sensor normalization)

  • Encoders x4 -Vision

    • vit-base16-embed-v1 (ViT-base/16 → 768d embedding)
    • resnet50-proj-v1 (ResNet50 + projector → 512d) -Audio
    • wav2vec2-base-embed-v1 (wav2vec2-base → 512d)
    • hubert-large-embed-v1 (HuBERT-large → 1024d) -Text
    • sbert-all-mpnet-v1 (SBERT / mpnet-base-v2 → 768d)

  • Spiking / Event

    • snn-dvs-embed-v1 (embedding for SNN/DVS)

  • Inference node (Inference x6)

  • LM (short text/dialogue)
    • gpt-small-v1 (GPT type small ~300M)
    • gpt-medium-v1 (GPT medium ~1.5B)
    • gpt-large-v1 (GPT type large ~6B) *Depends on demand -Classifier/Detector
    • yolox-s-intel-v1 (YOLOX small / detector)
    • fasterrcnn-res50-v1 (Faster-RCNN Res50)
  • Spiking-LM
    • spiking-lm-core-v1 (spiking generation model)

  • Ensemble / Multimodal

    • multimodal-ensemble-v1 (multimodal integration layer) -RAG-support
    • rag-lite-v1 (retriever + generation wrapper)

  • Decision node (Decision x2)

  • Planner
    • planner-rl-ppo-v1 (PPO base planner)

  • Controller

    • motor-controller-dnn-v1 (controller model)

  • Memory node (Memory x3)

  • Vector DB (separate from production artifacts: config / index templates)

    • milvus-schema-v1 (vector DB schema definition) -Episode storage
    • minio-log-schema-v1

  • Training node (Trainer x1)

  • trainer-ddp-manager-v1 (distributed learning job management)

  • Aggregation/arbitration node (Aggregator x2)

  • federator-agg-v1 (safety aggregation protocol)

  • result-aggregator-v1 (output aggregation, reliability evaluation)

  • Management/Utilities (Management x2)

  • auth-service-v1 (API key/RBAC service)
  • monitoring-stack-v1 (Prometheus/Grafana/ELK settings)
  • Format example: <component>-<base-model>-<purpose>-v<major>
  • Example: vision-vit-base16-embed-v1 → component=vision, base-model=vit-base16, purpose=embed, version=v1
  • Metadata to record (artifact manifest):
  • artifact_name, model_version, base_model, task, embedding_dim, quantized (bool), precision (fp32/fp16/int8), training_config_hash, train_data_tags, license, created_at, node_type, privacy_level

Notes (implementation specifications): - For generation scripts/training scripts, create artifact_manifest.json in the run save directory and include it in the upload ZIP. - Flag names used in CLI/front end (existing implementation): --artifact-name, --precision, --quantize (store_true), --privacy-level, --node-type. These are reflected in the manifest. - If artifact_name is not specified, it will be automatically generated and follow the recommended format prefix ({node_type}.{model_category}.{model_variant}.{run_name}.{timestamp}).

3. Frontend learning form parameters → artifact generation mapping

When triggering training on the frontend, it shows the key parameters the user will enter and which fields/settings will be reflected in the final generated artifact.

  • Input parameters (example):
  • component (selection): artifact_name prefix of the corresponding artifact (e.g. vision, audio, text, spiking)
  • base_model (selection/text): Pretrained base (e.g. vit-base16, wav2vec2-base, gpt-small-v1) → base_model meta
  • task (selection): embed / classification / lm-finetune / detection → task meta
  • embedding_dim (number): embedding dimension → embedding_dim
  • hidden_size, num_layers, num_heads (number): Architectural change → stored in model_config
  • max_seq_length / sample_rate / input_size: Model input/output specifications → input_spec
  • batch_size, learning_rate, optimizer, epochs, weight_decay: Training settings → training_config (and generate training_config_hash)
  • precision (selection): fp32/fp16/int8 → precision, quantized flag
  • quantize (bool): If True, perform quantization post processing within the job → append quantized=true to the artifact name (e.g. -int8)
  • checkpoint_interval (number): Checkpoint save frequency → checkpoint_policy
  • augmentations / preprocessing_profile: Data preprocessing → data_prep_profile
  • train_data_tags (tag list): Which dataset was used → train_data_tags meta

  • privacy_level (selection): none/dp/secure-agg → Apply differential privacy and secure aggregation to learning jobs

  • Mapping example (frontend input → generated artifact manifest):

  • component=vision, base_model=vit-base16, task=embed, embedding_dim=768, precision=fp16, quantize=false, batch_size=256, epochs=10 →

    • artifact_name: vision-vit-base16-embed-v1
    • manifest: {"base_model":"vit-base16","task":"embed","embedding_dim":768,"precision":"fp16","training_config_hash":""}

  • component=inference, base_model=gpt-small-v1, task=lm-finetune, max_seq_length=1024, learning_rate=2e-5, epochs=3, quantize=int8 →

    • artifact_name: gpt-small-v1-lm-finetune-int8-v1
    • manifest includes quantized:true, precision:int8, input_spec:{max_seq_length:1024}

4. Notes on implementation on the front end side (short)

  • When starting a training job, always calculate training_config_hash (JSON normalization → SHA256) and link it to the artifact. This allows reproducibility and comparison.
  • Quantization options can be selected in the UI by running a post-training-quantize step in the job or by selecting Quantization Awareness (QAT) during training.
  • Privacy settings (differential privacy and secure aggregation) are propagated to Trainer/Aggregator by putting privacy_level in the training job definition.

5. Next action suggestion

  1. Decide on a priority list of the above artifacts (first three) and automate the learning pipeline with CI. Recommended first three: vit-base16-embed-v1, wav2vec2-base-embed-v1, gpt-small-v1-lm-finetune-v1.
  2. Add the above parameters to the frontend learning form (frontend/pages/settings.py etc.) and design an API to submit the learning job via api_client.

File save location: docs/DISTRIBUTED_BRAIN_MODEL_ARTIFACTS.md