Mojana COMET - Active Inference + Stackelberg Game Theory

Fine-tuned COMET model with Active Inference and Stackelberg Game Theory reasoning for robotics applications.

Model Description

Based on Flan-T5-base (247M parameters), extended with:

  • Active Inference: xStateSpace, xPriorBeliefs, xEFEAnalysis, xSelectedAction, etc.
  • Stackelberg Game Theory: xGameAgents, xStackelbergEquilibrium, xCommitmentValue, etc.
  • Commonsense: ATOMIC relations (xWant, xNeed, xEffect, xIntent)
  • Physical Reasoning: PIQA (xPhysicalSolution)

Training Data

  • Active Inference examples: 1.87M
  • Stackelberg examples: 580K
  • PIQA: 2M
  • ATOMIC: 957K
  • Custom heuristics: 495K

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model = AutoModelForSeq2SeqLM.from_pretrained("Haya-as/mojana-comet-unified")
tokenizer = AutoTokenizer.from_pretrained("Haya-as/mojana-comet-unified")

# Active Inference query
prompt = "The robot detects smoke in the kitchen [xEFEAnalysis]"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Relations

Type Relations
Active Inference xStateSpace, xPriorBeliefs, xPriorEntropy, xGoalState, xActionSpace, xEFEAnalysis, xSelectedAction, xBeliefUpdateProcess
Stackelberg xGameAgents, xLeaderStrategies, xFollowerStrategies, xPayoffStructure, xFollowerBestResponses, xLeaderOptimization, xStackelbergEquilibrium, xCommitmentValue
ATOMIC xWant, xNeed, xEffect, xIntent, xAttr, HinderedBy, ObjectUse
Physical xPhysicalSolution, xPhysicalAlternative

License

MIT

Downloads last month
-
Safetensors
Model size
0.2B params
Tensor type
F32
·
Video Preview
loading

Datasets used to train Haya-as/mojana-comet-unified