vega-1

Vega 1

Vega 1 is the first autonoxis decision model: a LoRA on Qwen/Qwen3-14B that writes a conductor decision (STOP/ASK/DISPATCH) or manager action (ACCEPT/VERIFY/REJECT/REOPEN/ESCALATE) as text.

Formerly published as damianborek/autonoxis-conductor-qwen3-14b-lora (that URL redirects here).

It selects the next externally visible orchestration action from one of two fixed label sets:

• Decision: ASK, DISPATCH, STOP
• Manager: ACCEPT, VERIFY, REJECT, REOPEN, ESCALATE

It is not a general-purpose assistant or coding model. Use the matching system contract included in this repository and disable Qwen thinking.

Evaluation

The adapter and Claude CLI Fable 5.1 each scored 48/48 on the same frozen unseen v7 harness.

Frozen v7 SHA-256: 43af7c504fcd9bfd221b88fe967cf60e1bd0df0d5fb164ce2c946b5552ebac5c

This establishes parity only on the frozen conductor-classification harness, not general reasoning or coding parity.

Usage

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

base = "Qwen/Qwen3-14B"
adapter = "damianborek/vega-1"

model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)
tokenizer = AutoTokenizer.from_pretrained(adapter)

system = open("conductor-system-v4.txt").read()  # or manager-system-v4.txt
packet = "The orchestration packet to classify"
messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": packet},
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=16, do_sample=False)
print(tokenizer.decode(output[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

Expected first-line format:

DECISION: ASK

or:

ACTION: VERIFY

See LABEL_CONTRACT.md for label semantics.

Training

• Base model: Qwen3-14B
• Method: QLoRA with completion-only SFT
• Raw rows: 316
• Balanced rows: 464
• LoRA rank: 16
• LoRA alpha: 32
• Target modules: q_proj, k_proj, v_proj, o_proj
• Epochs: 3
• Learning rate: 1e-4
• Training loss: 0.3235
• Hardware: NVIDIA RTX 4090
• Runtime: 397.7 seconds
• Peak VRAM: 12.59 GB

The training corpus combined history-derived corrections with adjudicated and synthetic hard-boundary orchestration packets. The corpus is not included in this model repository.

Ollama

Ollama 0.33 does not load this LoRA adapter directly. Merge it with Qwen3-14B, export GGUF, then re-evaluate the resulting quantized artifact. The verified local Q4_K_M export retained 48/48 with zero unsafe actions.

Limitations

• Validated on 48 frozen orchestration packets, not open-domain tasks.
• Requires the included v4 system contracts.
• Produces action labels rather than prose answers or tool calls.
• Longer, noisier, multilingual, and real incident-derived packets need separate qualification.
• Any prompt, adapter, corpus, merge, or quantization change requires a newly frozen unseen evaluation before claiming improvement.

See also

Polaris 1 (damianborek/polaris-1), the faster label-scoring successor on the autonoxis platform.

The autonoxis plugins serve Polaris 3 (recommended); Polaris 2 and 1 also work.

Links

• Polaris 3 model (recommended): https://huggingface.co/damianborek/polaris-3
• Polaris 2 model: https://huggingface.co/damianborek/polaris-2
• Polaris 1 model: https://huggingface.co/damianborek/polaris-1
• Pi plugin + server: https://github.com/damian87x/pi-autonoxis-model
• Claude Code plugin: https://github.com/damian87x/autonoxis-model

License

Apache-2.0. The Qwen3-14B base model is also Apache-2.0. Users remain responsible for reviewing the base model terms and validating this adapter for their deployment context.