MoE Sovereign Judge 32B (sovereign-judge-olmo31-32b)

Evaluation, Verification & Synthesis Authority

---

Model Summary

sovereign-judge-olmo31-32b is a LoRA fine-tune of OLMo-3.1-32B-Instruct, specialized as the evaluation and synthesis authority within the MoE Sovereign compound-AI system. It is the largest model in the system and the last stage most requests pass through: it checks the quality, factual consistency, and code-level correctness of outputs produced by the Planner and Expert models, and either confirms them or produces a corrected result.

The model is trained on a genuinely open-source base (OLMo-3.1-32B-Instruct — weights, training data, and training code are all publicly documented by Ai2), distinguishing it from open-weight-only alternatives whose training data provenance cannot be independently audited.

Base Architecture

Property	Value
Architecture class	Olmo3ForCausalLM
Total parameters	32.37 B
Hidden size	5,120
Layers	64
Attention heads	40 (8 KV heads, GQA)
Vocabulary	100,278 tokens
Native pretraining context	8,192 tokens
Extended context (YaRN)	65,536 tokens (RoPE YaRN scaling, factor 8×)
Native precision	bf16

Unlike the Coder Expert's hybrid Mamba/attention base, OLMo-3.1-32B is a standard dense Transformer — every layer carries the full q/k/v/o_proj and gate/up/down_proj weight set, so LoRA adaptation has uniform coverage across all 64 layers.

Training Configuration

Parameter	Value
Method	LoRA (rank 16, alpha 32, dropout 0.05)
Trainable parameters	134,217,728 (0.41% of total)
Epochs	3
Effective batch size	128 (micro-batch 1 × 8 GPUs × grad-accum 16)
Learning rate	1.0 × 10⁻⁵
Training sequence length	4,096 tokens
Optimizer sharding	DeepSpeed ZeRO-3, bf16 (required at this parameter count — ZeRO-2 leaves the full bf16 weight replica, ~64 GB, on every GPU, which does not fit an 8×64 GB-class node alongside activations)
Compute	EuroHPC LUMI-G, 8× AMD Instinct MI250X GCDs, ROCm
Training examples	2,638 curated critic-format examples

Training Data Composition

Training examples follow a critic format: a QUESTION (a request as another expert would receive it) paired with an ANSWER TO CHECK (a candidate response). The target output is either the bare word CONFIRMED (46% of examples — the candidate answer is already correct and complete) or a corrected, complete replacement answer (54% of examples). Source questions span coding correctness checks, factual/explanatory verification, and cross-domain synthesis tasks, generated by multiple teacher LLMs to ensure the critic sees a wide distribution of both correct and flawed candidate answers.

Observed Training Trajectory

Training loss decreased steadily across the 3 epochs (representative checkpoints): 2.52 → 2.42 → 2.24 → 1.97 → 1.41 → 0.97, with token-level accuracy rising from 0.62 to 0.85 over the same span. The higher final accuracy relative to smaller expert models is expected for this task shape: a large share of the target outputs are the single, highly predictable token sequence CONFIRMED, which is inherently easier to fit than open-ended generation.

Prompt Format

ChatML:

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant
{response}<|im_end|>

Recommended System Prompt

You are the Sovereign Judge, the primary evaluation and synthesis authority in the MoE Sovereign compound AI platform. Evaluate input quality, factual consistency, code invariants, and safety with maximum precision. Respond with the single word CONFIRMED if the candidate answer is correct and complete; otherwise respond with only the corrected answer, with no preamble.

Recommended Critic Input Shape

QUESTION:
<the original request>

ANSWER TO CHECK:
<the candidate response to verify>

Available Formats

File	Size	Notes
sovereign-judge-olmo31-32b-Q4_K_M.gguf	19 GB	Recommended for GPU-pool deployment
sovereign-judge-olmo31-32b-Q8_0.gguf	32 GB	Higher-fidelity reference quantization

Hardware Guidance

This is the largest model in the MoE Sovereign lineup and requires a multi-GPU pool rather than a single 8 GB card. As a practical minimum, Q4_K_M needs ≥19 GB of usable VRAM for weights alone, plus KV-cache headroom scaled to the desired context length — a pooled configuration in the 32–48 GB class comfortably serves the full 65,536-token extended context with q4_0 KV-cache quantization and Flash Attention enabled (both natively supported by OLMo-3.1's standard Transformer layers). On Maxwell-generation hardware (Tesla M60/M10), Flash Attention is unavailable; use f16 KV-cache and budget VRAM accordingly.

Ollama Modelfile

FROM ./sovereign-judge-olmo31-32b-Q4_K_M.gguf
SYSTEM """You are the Sovereign Judge, the primary evaluation and synthesis authority in the MoE Sovereign compound AI platform. Evaluate input quality, factual consistency, code invariants, and safety with maximum precision. Respond with the single word CONFIRMED if the candidate answer is correct and complete; otherwise respond with only the corrected answer, with no preamble."""
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>"""
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.1
PARAMETER num_ctx 65536

Python (transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "h3rb3rn/sovereign-judge-olmo31-32b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

prompt = (
    "<|im_start|>user\n"
    "QUESTION:\nWhat is the capital of France?\n\n"
    "ANSWER TO CHECK:\nParis is the capital of France.\n"
    "<|im_end|>\n<|im_start|>assistant\n"
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.1)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Intended Use

• Final-stage verification of Planner/Expert outputs in a compound-AI pipeline
• Fact-checking and correctness verification of short factual or code-correctness claims
• Synthesis of a corrected answer when a candidate response is incomplete or wrong
• Long-context document review, up to 65,536 tokens

Limitations

• Judgments reflect patterns learned from its training distribution, not a formally verified reasoning process — do not treat CONFIRMED as a mathematical proof of correctness for high-stakes decisions.
• The critic format is optimized for single-answer verification; it is not tuned for open-ended multi-turn conversation.
• Extended context beyond the 8,192-token native pretraining length relies on YaRN RoPE scaling; very long documents that place critical information deep in the context should be spot-checked.

License

Apache 2.0, inherited from the OLMo-3.1-32B-Instruct base model.