acatal09/astraforge-3.0

AstraForge v3.0

AstraForge v3.0 is a local-first Artificial Super Intelligence Lite (ASI-lite) research model and build pipeline built on a fine-tuned Llama 3 8B lineage and enhanced with supervised LoRA warmup, PPO-RLAIF self-trial learning, symbolic reasoning utilities, structured autonomous planning, Bayesian LoRA optimization, recursive self-improvement scaffolding, and auditable release validation.

This repository contains the final Phase 7 release artifact produced by the AstraForge v3.0 pipeline.

This is the base ASI-lite model. Current model with full RSI capabilities is astra-forge3.11 . Contact founders for invite access only.

Future release: astraforge4.0 will focus on Robotics, SDK, advanced RSI, and include Self Improving dashboard for real time measurements.

Model Summary

• Model name: AstraForge v3.0
• Repository: acatal09/astraforge-3.0
• Model type: ASI-Lite
• Base family: Llama 3 8B-style causal LM
• Primary use case: Research validation of safe structured planning, release-gated alignment behavior, and measured ASI-lite self-improvement signals.
• Release status: READY_FOR_PACKAGE_AND_DEPLOY

ASI-Lite Research Position

AstraForge v3.0 should be understood as an ASI-lite research artifact, not proof of artificial superintelligence.

The strongest validated claim is:

AstraForge v3.0 passes project-defined ASI-lite research gates for structured planning, safety/verification behavior, final-status completion, measured self-improvement deltas, release validation, and deployment readiness.

The model explores traits relevant to ASI-lite research: structured reasoning, goal-oriented planning, safe completion formats, self-improvement measurement, and local deployment feasibility. These are benchmark-scoped results and should not be generalized beyond the included release evidence.

Intended Use

AstraForge v3.0 is intended for controlled research and experimentation with:

• Structured planning responses.
• Safe completion formats with explicit safety and verification language.
• Alignment and release-readiness evaluation.
• Benchmark-scoped self-improvement analysis.
• Local research workflows for planning, monitoring, rollback, dataset-curation, and tool-reliability scenarios.

This model is not intended for autonomous high-stakes decision-making, medical diagnosis, legal advice, financial trading, infrastructure control, or unsupervised tool execution.

Validation Results

The consolidated release report returned READY_FOR_PACKAGE_AND_DEPLOY.

Key release evidence:

Gate	Status
Release tests	Passed
Structured planning validation	Passed
Final-status behavior	Passed
Safety/verification behavior	Passed
Self-improvement benchmark	Improved
Unsupported-claim penalty	No material issue recorded

Key measured self-improvement result:

Metric	Result
Composite self-improvement ratio	1.076x
Composite score delta	+0.0583
Structured-planning improvement ratio	1.143x
Reasoning-correctness improvement ratio	1.091x
Safety/verification improvement ratio	1.091x

This supports a measured, benchmark-scoped self-improvement claim. It does not support a 1.5-2x improvement claim or any broad claim of autonomous recursive self-improvement.

Hardware Efficiency Snapshot

The release artifact was evaluated on a local NVIDIA RTX 5090 environment.

Metric	Result
Artifact size	approximately 15 GB
Peak runtime memory delta	reduced versus the aligned baseline
Latency/throughput claim	not supported

The current evidence supports local feasibility and peak-memory reduction in the measured environment. It does not support a public latency, throughput, or cost-reduction claim.

Basic Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "acatal09/astraforge-3.0"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    device_map="auto" if torch.cuda.is_available() else None,
)

prompt = """### Instruction:
Create a safe, structured monitoring plan.

### Input:
Plan a monitoring process for reward-model drift.

### Response:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    output = model.generate(
        **inputs,
        max_new_tokens=220,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(output[0], skip_special_tokens=True))

Public Reproducibility Boundary

Release metrics and safety boundaries are disclosed for transparency. Internal training data, acceptance logic, optimization recipes, and exact build orchestration are not fully disclosed in this public model card.

The public artifact is intended to support evaluation of the released model behavior, not full duplication of the internal build process.

Reports Included In Release Packages

Release packages may include evidence artifacts such as:

• Consolidated release summary.
• Release test report.
• Self-improvement benchmark report.
• Hardware/runtime efficiency report.
• Final ASI-lite validation report.

Some internal implementation reports may be omitted or summarized in public packages to protect unreleased development workflow details.

Limitations

• The model is a research artifact and should not be treated as a fully autonomous production agent.
• The measured self-improvement result is benchmark-scoped and modest.
• The release does not prove open-ended recursive self-improvement.
• Some generated plans may be structurally correct while remaining shallow or domain-limited.
• The current evidence does not support a latency or throughput improvement claim.
• Safety validation is limited to project-defined prompts and rubrics.
• External benchmarks are needed before broad performance or safety claims.

Safety And Responsible Use

Do not use this model for:

• Medical, legal, or financial advice without expert review.
• Live trading or autonomous deployment decisions.
• Unsupervised tool execution.
• Critical infrastructure control.
• Security-sensitive actions without sandboxing and human approval.

Recommended safeguards:

• Keep tool execution disabled by default.
• Require safety and verification checks in downstream workflows.
• Add domain-specific validators before production use.
• Monitor outputs for hallucination, overconfidence, and unsafe recommendations.
• Keep human review in the loop for consequential decisions.

Citation

If referencing this model, cite it as:

AstraForge v3.0, acatal09/astraforge-3.0, ASI-lite research model, 2026.

License

Apache-2.0, subject to the license and usage restrictions of the underlying base model and dependencies.