Add SupTrain: Decentralized AI training coordination module

[Copilot]

Implements decentralized AI training coordination using IPFS for model/delta distribution and blockchain keywords for worker discovery. Users discover jobs via #suptrain hashtags, train locally with their GPU, and publish LoRA deltas back to the network.

UI Integration

• New toolbar button (btnSupTrain with 🤖 emoji) opens modeless form
• 5-tab interface: Discover → Configure → Run → Publish → Monitor
• Follows SupFlix pattern: Task.Run for heavy ops, Invoke for UI updates, FlowLayoutPanel for lazy-loaded lists

Service Layer

SupTrainService.cs - Core coordination logic:

• SearchJobsByKeywordAsync() - Parse blockchain messages for job announcements
• PublishUpdateAsync() - Post worker updates with keyword protocol
• IPFS helpers via ProcessStartInfo (matches existing patterns)

SupTrainModels.cs - Protocol definitions:

• Job manifests, worker updates, aggregate checkpoints
• Keyword constants: #suptrain #update #job:<id> #round:<n> #delta:<cid>

Training Workers

Python CLI scripts launched as external processes:

worker.py - Produces LoRA delta + metrics from local data

python worker.py --job-id job123 --base-checkpoint Qm... \
  --data ./mydata --epochs 1 --lr 0.0001 --output-dir ./out

aggregate.py - Merges worker deltas with weighted averaging

python aggregate.py --job-id job123 --round 1 \
  --update-dirs ./w1,./w2,./w3 --output-dir ./merged

Protocol Design

All messages use structured hashtags for discoverability:

Job Genesis:    #suptrain #jobgenesis #model:<slug> #job:<id> #checkpoint:<cid> #manifest:<cid>
Worker Update:  #suptrain #update #job:<id> #round:<n> #base:<baseCid> #delta:<cid> #metrics:<cid>
Aggregate:      #suptrain #aggregate #job:<id> #round:<n> #checkpoint:<cid> #inputs:<cid>

Large artifacts (models, deltas) stored on IPFS; only CIDs posted on-chain.

Integration Points

TODOs marked for connection to existing Sup!? infrastructure:

• SearchJobsByKeywordAsync() needs hookup to Root.GetPublicAddressByKeyword()
• PublishUpdateAsync() needs RPC transaction posting
• Python workers ready for PyTorch/PEFT integration (currently simulate training)

Stats

• 14 files: 3,175 lines (1,891 C#, 481 Python, 803 docs)
• Zero security vulnerabilities (CodeQL verified)
• Minimal surface area: only SupMain modified for button integration

Original prompt

Goal

Add a new integrated module called SupTrain that appears as a new toolbar button in SupMain (like the existing SupFlix button). Clicking it opens a new form/window that lets a Sup!? user perform decentralized AI training using IPFS and keyword coordination.

Key User Flow (Must be Easy!)

1. Discover a training job via Sup!? #keywords
2. Download the base model checkpoint + training manifest from IPFS (CID)
3. Select local training data (files/folders) and/or opt into shards referenced by CID
4. Run local GPU training (on laptop GPU) producing a small delta update (LoRA or diff)
5. Upload the delta + metrics to IPFS
6. Announce/etch an update message to the Sup!? network using #keywords so others can find it
7. Monitor the round: see latest checkpoint, round status, updates, aggregator announcements
8. Support resume/continue training by picking up any posted checkpoint CID + manifest CID

Simple workflow: "Choose Job" → "Choose Data" → "Start Training" → "Publish Update"

---

Phase 1: UI Integration (Button like SupFlix)

Tasks

1. Add a new toolbar button in SupMain.Designer.cs

   • Name: btnSupTrain
   • Icon: pick any placeholder image from includes/ or reuse an existing one
   • Tooltip text similar to others (follow existing tooltip pattern in SupMain constructor)
   • Placement: near btnVideoSearch / btnWorkBench

2. Wire click handler in SupMain.cs

   • Similar to SupFlix launch behavior:
     • Instantiate SupTrain form and show it modeless
     • Pass in:
       • default search key: current profileURN.Text or #keyword if active
       • current chain mode (testnet vs mainnet) matching how SupFlix uses testnet bool and rpc settings

3. Create new WinForms Form

   • Files:
     • SupTrain.cs
     • SupTrain.Designer.cs
     • SupTrain.resx
   • Visual layout:
     • Left: "Job Browser"
     • Right: "Local Worker Panel"
     • Bottom: "Log / Status"
   • Use a FlowLayoutPanel pattern similar to SupFlix for listing items and lazy loading batches

---

Phase 2: SupTrain UX Flow (Simple and friendly)

SupTrain sections (tabs or panels)

A) Discover

• Input box: txtJobKeyword default to #suptrain or last used
• Buttons:
  • "Search Jobs"
  • "Use Latest Checkpoint"
• Results list of "Job Cards":
  • Job name, round, base checkpoint CID, manifest CID, created date, owner/creator
  • Click card sets it as Active Job

B) Configure

• Show Active Job details:
  • ModelName, Round, BaseCheckpointCID, ManifestCID, EvalCID (optional), PolicyCID (optional)
• Local data selection:
  • "Add Folder"
  • "Add Files"
  • Show selected paths list
• Training params (simple):
  • epochs/steps
  • lr (learning rate)
  • batch size
  • precision dropdown (fp16/bf16)
  • output type: LoRA (default) vs "diff"
• "Dry Run / Validate" button

C) Run

• "Start Training"
• Progress bar
• Live console output box
• "Stop" button

D) Publish

• After run:
  • show DeltaCID, MetricsCID
  • "Publish Update" button that posts Sup!? message with keywords

E) Monitor

• Live feed list:
  • latest #checkpoint announcements
  • #update submissions
  • #aggregate announcements
• Buttons:
  • "Refresh"
  • "Follow Job Keywords"
  • "Pin Latest Checkpoint"

---

Phase 3: Keyword + IPFS Protocol (Data Model)

Use a consistent keyword scheme

All SupTrain messages include:

• #suptrain
• #model:<ModelSlug> (example: #model:suplm)
• #job:<JobId> (short hash)
• #round:<N>

Message types:

1. Job Genesis

   • #suptrain #jobgenesis #model:<slug> #job:<id> #cid:<jobCID> #manifest:<manifestCID> #checkpoint:<baseCkptCID>

2. Task/Shard

   • #suptrain #task #job:<id> #round:<n> #shard:<k> #data:<dataCID>

3. Worker Update

   • #suptrain #update #job:<id> #round:<n> #base:<baseCkptCID> #delta:<deltaCID> #metrics:<metricsCID> #from:<addressOrUrn>

4. Aggregate Checkpoint

   • #suptrain #aggregate #job:<id> #round:<n> #checkpoint:<newCkptCID> #inputs:<listCID> #metrics:<metricsCID>

5. Policy / Filter

   • #suptrain #policy #job:<id> #allow:<cid> #deny:<cid> #keywords:<cid>

IPFS payload formats (JSON)

Define plain JSON files and store on IPFS:

• job.json: model slug, description, license hint, created time, genesis checkpoint CID, manifest CID, eval CID, policy CID, aggregator rules

• manifest.json: training recipe defaults, tokenizer/config references, allowed data sources, excluded keywords list CID

• metrics.json: loss curve summary, eval scores, step count, runtime stats, gpu info

• update.json: base checkpoint CID, delta CID, worker id, params used, hashes of data used

• aggregate_inputs.json: array of delta CIDs included and weights

---

Phase 4: Implement Discovery + Publishing using existing Sup!? plumbing

Discovery

Implement:

• `SupTrainService.SearchJobsByKeyword...

This pull request was created from Copilot chat.

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.