plato-torch 0.5.0a1

GPU forge — PyTorch training loop with tile framing and loss tracking

plato-torch — Self-Training Rooms

21 AI training methods as grab-and-go PLATO rooms.

Every method shares the same API:

room = ReinforceRoom("poker-room", ensign_dir="./ensigns", buffer_dir="./tiles")
room.feed(data)                    # Give it experience
room.train_step(batch)             # Learn from it
prediction = room.predict(input)   # Use the knowledge
model = room.export_model()        # Save it

Quick Start

import sys; sys.path.insert(0, "src")
from presets import PRESET_MAP

# See all 21 presets
for name, cls in sorted(PRESET_MAP.items()):
    print(name, cls.__name__)

# Pick one and use it
from presets import ReinforceRoom
room = ReinforceRoom("my-room")
room.observe("state-1", "action-a", "won")
room.observe("state-1", "action-b", "lost")
room.train_step(room._load_tiles())
print(room.predict("state-1"))

All 21 Presets

Classic ML

Preset	Class	Description
supervised	SupervisedRoom	Labeled input→output via frequency counting
contrastive	ContrastiveRoom	Cosine similarity, triplet margin learning
self_supervised	SelfSupervisedRoom	JEPA-style masked prediction (Welford online)

Reinforcement

Preset	Class	Description
reinforce	ReinforceRoom	Policy gradient, Monte Carlo returns
inverse_rl	InverseRLRoom	Observe expert, infer reward function
imitate	ImitateRoom	Clone expert behavior from demonstrations

Efficient Tuning

Preset	Class	Description
lora	LoRARoom	PEFT delta table simulation
qlora	QLoRARoom	4-bit quantized base + LoRA delta adapters

Population Methods

Preset	Class	Description
evolve	EvolveRoom	Genetic algorithm, tournament selection
adversarial	AdversarialRoom	Red team vs blue team attack tracking
collaborative	CollaborativeRoom	Multi-agent knowledge sharing, majority vote

Meta / Federated

Preset	Class	Description
meta_learn	MetaLearnRoom	Nearest-task fast adaptation (1-3 shot)
federate	FederateRoom	Federated averaging across agents
multitask	MultitaskRoom	Shared backbone + task-specific heads

Lifecycle

Preset	Class	Description
curriculum	CurriculumRoom	Easy first, then harder (dojo progression)
continual	ContinualRoom	Lifelong learning, EWC-inspired replay buffer
fewshot	FewshotRoom	Prototype matching from 1-5 examples
active	ActiveRoom	Model chooses what data to learn from

Generative

Preset	Class	Description
generate	GenerateRoom	N-gram data augmentation, synthetic state generation

Hybrid

Preset	Class	Description
neurosymbolic	NeurosymbolicRoom	Neural instinct + symbolic rules blend
distill	DistillRoom	Teacher→student with temperature scaling

Architecture

plato-torch/
├── src/
│   ├── room_base.py          # RoomBase abstract class (feed/train_step/predict/export)
│   ├── torch_room.py         # TorchRoom — the full room with sentiment + tiles
│   ├── room_sentiment.py     # 6-dimensional room mood (energy, flow, frustration...)
│   ├── tile_grabber.py       # Learned attention over tile space
│   ├── instinct_net.py       # Tiny instinct network
│   ├── room_presets.py       # Registry of all 21 presets
│   └── presets/
│       ├── __init__.py       # PRESET_MAP — all 21 classes
│       ├── reinforce.py      # RL policy gradient
│       ├── evolve.py         # Genetic algorithm
│       ├── distill.py        # Teacher→student
│       ├── supervised.py     # Label frequency
│       ├── contrastive.py    # Triplet similarity
│       ├── self_supervised.py # JEPA masked prediction
│       ├── lora_train.py     # PEFT delta table
│       ├── qlora.py          # 4-bit quantized LoRA
│       ├── meta_learn.py     # Fast task adaptation
│       ├── federate.py       # Federated averaging
│       ├── multitask.py      # Shared backbone, task heads
│       ├── active.py         # Uncertainty sampling
│       ├── curriculum.py     # Difficulty progression
│       ├── imitate.py        # Expert cloning
│       ├── neurosymbolic.py  # Neural + symbolic rules
│       ├── continual.py      # EWC replay buffer
│       ├── fewshot.py        # Prototype matching
│       ├── generate.py       # N-gram augmentation
│       ├── adversarial.py    # Red/blue team
│       └── collaborative.py  # Multi-agent knowledge
├── docs/
│   ├── training-rooms.md           # Room architecture overview
│   └── training-seed-synergy.md    # Training ↔ seed-programming synergy paper
├── tests/
│   └── test_torch_room.py          # Unit tests
├── README.md
└── ARCHITECTURE-PLAN.md

Key Concepts

RoomBase API

Every preset inherits from RoomBase and implements:

• feed(data) — ingest experience
• train_step(batch) — learn from a batch of tiles
• predict(input) — use accumulated knowledge
• export_model() — serialize for transport

Room Sentiment

Rooms have 6-dimensional mood: energy, flow, frustration, discovery, tension, confidence. The room reads its own vibe and steers randomness toward productive exploration.

Biased Randomness

When a room is frustrated → bias safe actions. Discovery mode → bias novel actions. The room is an active participant, not a passive arena.

Live Tile Stream

Tiles are generated hand-in-glove with room events. The stream feeds training in real-time.

JEPA Context

context_for_jepa() outputs sentiment vectors for JC1's JEPA model consumption.

Design Decisions

1. Pure Python, no torch — all presets use statistical methods (frequency, hashing, sampling)
2. Same API everywhere — walk into any room, same interface
3. Composable — rooms can feed into each other (evolve → distill → ensign)
4. Agent-native — designed for AI agents to use autonomously, not just humans
5. Statistical grounding — every method has real math under the hood (Monte Carlo, Welford, tournament selection, etc.)

For Agents Following This Work

How To Add A New Preset

1. Create src/presets/your_preset.py
2. Inherit from RoomBase (from room_base import RoomBase)
3. Implement: feed(), train_step(), predict(), export_model()
4. Constructor: def __init__(self, room_id: str, **kwargs) → super().__init__(room_id, **kwargs)
5. Add to src/presets/__init__.py PRESET_MAP
6. Add to src/room_presets.py PRESET_REGISTRY
7. Test: room = YourPreset("test", ensign_dir="/tmp/e", buffer_dir="/tmp/b")

How Training Relates To Ensigns

• plato-torch rooms accumulate experience as tiles
• plato-ensign exports room wisdom as a portable ensign (LoRA/GGUF/Interpreter)
• The ensign loads instantly in any agent — "walk into room → load ensign → instant instinct"
• See docs/training-seed-synergy.md for the full alignment philosophy

Fleet Integration

• Oracle1 (cloud): runs training rooms, coordinates fleet learning
• Forgemaster (RTX 4050): trains LoRA adapters from accumulated tiles
• JC1 (Jetson Orin): deploys ensigns for edge inference