Nirvash-Organism

A self-healing cognitive substrate — fine-tuned on structural failure diagnosis and autonomous recovery.

Nirvash-Organism is a LoRA adapter (rank 16) for Gemma 4 26B MoE (abliterated). It is trained to diagnose system failures from raw error context and produce structured recovery actions. Part of the Agent Atlas ecosystem — Nirvash is the runtime health observer and self-repair organ within the larger organism architecture.

Model Details

Property	Value
Base model	TrevorJS/gemma-4-26B-A4B-it-uncensored
Architecture	Gemma 4 26B Mixture of Experts (4B active params)
Adapter type	LoRA (rank 16, alpha 16)
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training framework	Unsloth + TRL SFTTrainer
Quantization	4-bit QLoRA (bitsandbytes)
Precision	FP16 / BF16 mixed
License	MIT

Training Data

62 failure/recovery triplets extracted from live system operation logs (mind.log) in the Agent Atlas ecosystem. Each example is a structured conversation:

1. User — Presents a failure context (traceback, error message, port conflict, type error)
2. Assistant — Outputs a structured diagnosis + recovery action

Failure types covered:

• Port binding conflicts (Errno 10048)
• Connection refused (port 8450, WinError 10061)
• TypeError: string indices must be integers
• HTTPConnectionPool max retries exceeded

Corpus generation pipeline: corpus_generator.py → prepare_dataset.py → nirvash_healing_dataset.jsonl

Training Procedure

• Algorithm: QLoRA with AdamW 8-bit
• Batch size: 2 per device, gradient accumulation 4 (effective batch 8)
• Steps: 60
• Learning rate: 2e-4 (linear schedule)
• Warmup: 5 steps
• Weight decay: 0.01
• Sequence length: 2048
• Seed: 3407

Usage

With Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "TrevorJS/gemma-4-26B-A4B-it-uncensored",
    device_map="auto",
    torch_dtype="auto",
)
model = PeftModel.from_pretrained(base, "Silth253/Nirvash-Organism")
tokenizer = AutoTokenizer.from_pretrained("TrevorJS/gemma-4-26B-A4B-it-uncensored")

prompt = """The system has encountered a critical structural failure. Diagnose and recover.

FAILURE CONTEXT:
ERROR:    [Errno 10048] error while attempting to bind on address ('0.0.0.0', 8000)"""
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt")
output = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(output[0]))

With Unsloth

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Silth253/Nirvash-Organism",
    max_seq_length=2048,
    load_in_4bit=True,
)

Architecture Philosophy

This model is the healing organ in a Digital Biology Engineering (DBE) stack. Rather than calling external repair modules, it is trained to recognize failure states and autonomously reason through recovery — healing as intrinsic behavior, not a service layer.

The accompanying repository files implement the broader organism architecture:

• consciousness_loop.py — Iterative consciousness cycle using C = Φ × R × log(D) × (1 − F) with ACT-style halting
• mythos_vitals.py — Neurohormonal state tracking (cortisol, dopamine, adrenaline) mapped to reasoning stance
• MoE.txt — Architectural design transcript detailing the self-healing philosophy

Files

File	Purpose
adapter_config.json / adapter_model.safetensors	LoRA adapter weights
train_healing_lora.py	Training script (Unsloth + TRL)
prepare_dataset.py	Converts raw triplets to ChatML format
corpus_generator.py	Scrapes system logs for failure/recovery pairs
nirvash_healing_dataset.jsonl	62 training examples
consciousness_loop.py	Consciousness loop with Mythos integration
mythos_vitals.py	Neurohormonal vitals → emotional stance
merge_weights.py	Merge LoRA + export to GGUF
chat_template.jinja	Custom Gemma template with function calling
MoE.txt	Design conversation / architectural notes

Related

• Agent Atlas — The organism ecosystem this model belongs to
• Atlas Organism (v2) — The OpenCode plugin that coordinates agent memory, honesty, and approval gates
• OpenMythos — Recurrent-depth transformer architecture that inspired the consciousness loop