atlas-1-lora

ATLAS-1: Neural World Model

ATLAS-1 is a LoRA adapter for Qwen3.5-35B-A3B (MoE: 35B total params, 3B active) that predicts execution outcomes before tasks run — a neural world model for AI orchestration.

What it does

Given a task description and system state, ATLAS-1 predicts:

• Success probability (0-1)
• Estimated duration (milliseconds)
• Token usage and cost (USD)
• Recommended agent and model
• Risk factors (specific, actionable warnings)
• Prediction confidence (0-1)

Example

Input (system state):

{
  "task": "Refactor the authentication service to use dependency injection",
  "taskType": "code",
  "complexity": "complex",
  "selectedAgent": "economist",
  "selectedModel": "deepseek-chat",
  "sessionLength": 45,
  "systemLoad": 0.82,
  "recentAgentSuccess": 0.6
}

Output (prediction):

{
  "willSucceed": 0.31,
  "estimatedDuration": 185000,
  "estimatedTokens": 22000,
  "estimatedCost": 0.0088,
  "recommendedAgent": "nova",
  "recommendedModel": "claude-sonnet-4-20250514",
  "riskFactors": [
    "Agent-task mismatch: economist for code refactoring",
    "High system load may cause timeouts",
    "Long session context adds noise"
  ],
  "confidence": 0.88
}

Training Details

• Base model: Qwen3.5-35B-A3B (hybrid linear attention + Mamba SSM + 256 MoE experts)
• Method: LoRA (r=32, alpha=64, targets: q_proj + v_proj)
• Training data: 10,500 examples (500 real execution outcomes + 10,000 distilled from Grok 4.1 Fast Reasoning)
• Epochs: 3 (1,770 steps)
• Best validation loss: 0.1952
• Training time: 22.0 hours on NVIDIA H100 80GB
• Precision: BF16 with BitsAndBytes 4-bit quantization (NF4)
• Optimizer: AdamW 8-bit, LR=2e-5 with cosine decay

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import json

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-35B-A3B", trust_remote_code=True)
model = PeftModel.from_pretrained(base_model, "hyperspaceai/atlas-1-lora")
tokenizer = AutoTokenizer.from_pretrained("hyperspaceai/atlas-1-lora")

system = "You are ATLAS-1, a neural world model for the Thor AI orchestration system. Given a structured description of a task and system state, predict the execution outcome. Respond with valid JSON only."
state = json.dumps({"task": "Fix the login bug", "taskType": "debug", "complexity": "moderate", "selectedAgent": "nova", "selectedModel": "grok-4-1-fast-reasoning", "sessionLength": 5, "systemLoad": 0.3, "recentAgentSuccess": 0.85})

prompt = f"<|im_start|>system\n{system}<|im_end|>\n<|im_start|>user\n{state}<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
prediction = json.loads(tokenizer.decode(outputs[0], skip_special_tokens=True).split("assistant\n")[-1])

Data Generation

Training data was generated using knowledge distillation from Grok 4.1 Fast Reasoning, which reasons about agent-task fit, model-task fit, system load effects, and realistic cost/duration estimates. Real execution outcomes from the Thor production system provide ground truth grounding.

Part of the Thor Ecosystem

ATLAS-1 runs as a pre-execution prediction step in the Thor AI orchestration pipeline, enabling proactive risk detection and resource optimization before tasks execute.

Developed by: HyperspaceAI