PEFT / LoRA adapter on IBM Granite 4.1-8B-Instruct (served here via unsloth/granite-4.1-8b), specialized for ChaosNexus Anvil Rhai plugin authorship, host APIs, and local MCP workflows.

Field	Value
Release name	ChaosNexus Tuned v1
Hub id	`TunedChaos/ChaosNexus_Tuned_v1`
Adapter type	LoRA (`r=32`, `alpha=64`, dropout `0.05`)
Train recipe	Continual PEFT from iter-6 → iter-8 goldens
Local checkpoint	`~/.unsloth/studio/outputs/ChaosNexus_Tuned_v1` (alias of iter-8)

Intended use

Local assistant for writing and repairing Rhai plugins that run inside ChaosNexus Anvil
Structured outputs compatible with ChaosNexus Forge approval (HITL)
Offline / air-gapped developer machines under sandbox + human-in-the-loop

Out of scope

Autonomous shell/network agents without ChaosNexus sandbox + human approval
General-purpose chat without ChaosNexus tooling context
Guaranteeing exploit-free behavior; always run plugins through Forge approval

Training data

Curated ShareGPT-style goldens in chaosnexus-tuned (datasets/iter8/, injectors under scripts/injectors/). Focus: real Anvil host signatures, TOML allowlists, MCP mesh hops, deny-by-default security. No Codex RAG in this train loop.

Evaluation (primary - publish this)

ChaosNexus Anvil golden rubric (18 prompts in chaosnexus-tuned/tests/Questions.md).

Scoring: Pass=1.0 / Partial=0.5 / Fail=0.0. Decode: greedy. Smoke gate: prompts 1 / 4 / 5 / 7 must be Pass or Partial.

Release	Mean	Smoke	Notes
ChaosNexus Tuned v1 (iter-8)	0.944	CLEAR	Full-version gate (≥0.90)
Alpha baseline (iter-6)	0.833	CLEAR	Alpha gate (≥0.70)

Full per-prompt table: chaosnexus-tuned/tests/eval_scores_v1.md (alias of iter-8 scores).
Raw generations: tests/eval_results_iter8.md, tests/eval_results_iter8_smoke.md.

Reproduce:

bash

cd chaosnexus-tuned
HIP_VISIBLE_DEVICES=0 ~/.unsloth/studio/unsloth_studio/bin/python scripts/run_evals.py \
  --model ~/.unsloth/studio/outputs/ChaosNexus_Tuned_v1 \
  --max_tokens 1024 --greedy \
  --output tests/eval_results_v1.md

See BENCHMARKS.md in this folder for Anvil + optional Open-LLM-style harness notes.

How to load

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "unsloth/granite-4.1-8b"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "TunedChaos/ChaosNexus_Tuned_v1")

Local Crucible (GGUF)

For ChaosNexus Crucible / Forge default download, use the quantized GGUF sibling:

TunedChaos/ChaosNexus_Tuned_v1-GGUF (Q4_K_M).

License

Base model: follow IBM Granite / Unsloth redistributor terms on the Hub.
Adapter / ChaosNexus packaging: AGPL-3.0-or-later unless a commercial license is obtained from Tuned Chaos.

Citation / links

Docs: https://chaosnexus.ai
Tuned repo: https://codeberg.org/TunedChaos/chaosnexus-tuned
Project: Tuned Chaos

TunedChaos/ChaosNexus_Tuned_v1

ChaosNexus Tuned v1 (ChaosNexus_Tuned_v1)