acatal09/astraforge-3.0

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:May 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AstraForge v3.0 by acatal09 is an 8 billion parameter ASI-lite research model built on a fine-tuned Llama 3 lineage with an 8192-token context length. It incorporates supervised LoRA warmup, PPO-RLAIF self-trial learning, and recursive self-improvement scaffolding. This model is primarily designed for research validation of safe structured planning, release-gated alignment behavior, and measured ASI-lite self-improvement signals.

Loading preview...

AstraForge v3.0: An ASI-Lite Research Model

AstraForge v3.0, developed by acatal09, is an 8 billion parameter research model built upon a fine-tuned Llama 3 lineage. It's characterized as an "ASI-lite" artifact, focusing on exploring traits relevant to artificial superintelligence research within controlled environments. The model integrates advanced training techniques including supervised LoRA warmup, PPO-RLAIF self-trial learning, symbolic reasoning utilities, structured autonomous planning, Bayesian LoRA optimization, and recursive self-improvement scaffolding.

Key Capabilities & Research Focus

  • Structured Planning: Excels at generating structured planning responses.
  • Safety & Verification: Designed for safe completion formats with explicit safety and verification language.
  • Measured Self-Improvement: Demonstrates a benchmark-scoped self-improvement ratio of 1.076x composite, with 1.143x in structured planning and 1.091x in reasoning correctness.
  • Local Deployment Feasibility: Evaluated for local feasibility and peak-memory reduction on an NVIDIA RTX 5090 environment.
  • Alignment Evaluation: Intended for evaluating alignment and release-readiness.

Intended Use Cases

AstraForge v3.0 is specifically intended for controlled research and experimentation in areas such as:

  • Developing and validating structured planning responses.
  • Evaluating safe completion formats and alignment behaviors.
  • Analyzing benchmark-scoped self-improvement.
  • Supporting local research workflows for planning, monitoring, rollback, dataset curation, and tool reliability scenarios.

It is crucial to note that this model is a research artifact and not intended for autonomous high-stakes decision-making, medical diagnosis, legal advice, financial trading, infrastructure control, or unsupervised tool execution.