kai-os

gemma4-31b-Opus-4.6-reasoning

Deploy Dedicated

README

License: apache-2.0

Base Model

Base model: google/gemma-4-31B-it
Adapter type: LoRA / QLoRA (peft)
Quantization: 4-bit NF4
Precision: BF16 compute

Dataset

Source dataset:

Crownelius/Opus-4.6-Reasoning-2100x-formatted

Local filtering applied before training:

Removed duplicate user prompts
Removed obviously bad prompt families and formatting noise
Kept reasoning-style rows only

Final local dataset stats:

Rows in source: 2160
Rows kept: 2025
Train rows: 1924
Validation rows: 101
Category mix: 1899 math, 126 code

Training Setup

Max sequence length: 4096
Epochs: 2
Learning rate: 1e-4
Per-device batch size: 1
Gradient accumulation: 8
Hardware: NVIDIA GH200

LoRA target modules were adapted for Gemma 4 wrapped linear layers:

q_proj.linear
k_proj.linear
v_proj.linear
o_proj.linear
gate_proj.linear
up_proj.linear
down_proj.linear

Validation Metrics

Final metrics from the completed run:

Eval loss: 3.6018
Eval perplexity: 36.66
Train runtime: 3723s
Epochs completed: 2.0

Published Base-Model Reference Benchmarks

The table below is included for context and comes from Google's official Gemma 4 31B Instruct model card. These are published base-model reference scores for google/gemma-4-31B-it, not adapter-specific evaluation results for this repository.

Table with columns: Benchmark, Gemma 4 31B, Gemma 3 27B (no think)
Benchmark	Gemma 4 31B	Gemma 3 27B (no think)
MMLU-Pro	85.2%	67.6%
AIME 2026 no tools	89.2%	20.8%
LiveCodeBench v6	80.0%	29.1%
Codeforces Elo	2150	110
GPQA Diamond	84.3%	42.4%
Tau2 (average over 3)	76.9%

Source:

google/gemma-4-31B-it

Usage

This repository contains a PEFT adapter, not a fully merged standalone model.

Load it with:

Base model: google/gemma-4-31B-it
Adapter: this repository

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

base_id = "google/gemma-4-31B-it"
adapter_id = "kai-os/gemma4-opus-reasoning-adapter-v1"

bnb = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    device_map="auto",
    quantization_config=bnb,
    torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(base, adapter_id)

Notes

This is a reasoning-focused adapter, not a benchmark-optimized release.
The benchmark table above is for the published base model, not this adapter.
It is best treated as an experimental distilled reasoning adapter.

Acknowledgements

Google for Gemma 4
The Opus reasoning dataset authors and maintainers
Hugging Face transformers, peft, and datasets

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

kai-os

Model Tree

Base

google/gemma-4-31B-it

Adapter

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality