This repository contains a LoRA adapter only for Oka on top of google/gemma-3-27b-it. It is an early S0 quality-improvement checkpoint trained from curated Hugging Face instruction, preference, math, and Japanese instruction rows ingested into RisingWave.

This is not a standalone model. You need access to google/gemma-3-27b-it and must comply with the base model license and access requirements.

Training

Base model: google/gemma-3-27b-it
Runner: Unsloth 4-bit
Adapter: LoRA r=16, alpha=32
Sequence length: 512
Batch: 1, gradient accumulation 4
Training rows fetched from RW: 1,200
Run: oka-unsloth-gemma3-27b-hf-quality-sft-100step-20260509

json

{
  "run_id": "oka-unsloth-gemma3-27b-hf-quality-sft-100step-20260509",
  "runner": "unsloth",
  "model_id": "google/gemma-3-27b-it",
  "strategy": "27b_hf_quality_sft",
  "dataset_labels": [
    "hf:allenai/tulu-3-sft-mixture/train",
    "hf:HuggingFaceH4/ultrachat_200k/train_sft",
    "hf:HuggingFaceH4/ultrafeedback_binarized/train_prefs",
    "hf:nvidia/OpenMathInstruct-2/train_1M",
    "hf:microsoft/orca-math-word-problems-200k/train",
    "hf:llm-jp/databricks-dolly-15k-ja/train"
  ],
  "fetched_rows": 1200,
  "training_examples": 1200,
  "output_dir": "/workspace/oka-unsloth-gemma3-27b-hf-quality-sft-100step-20260509",
  "train_runtime_sec_total": 484.53579926490784,
  "metrics": {
    "train_runtime": 348.3248,
    "train_samples_per_second": 1.148,
    "train_steps_per_second": 0.287,
    "total_flos": 3.21104786030592e+16,
    "train_loss": 1.297047061920166,
    "epoch": 0.3333333333333333
  },
  "max_memory_allocated_gb": 25.574787139892578,
  "max_memory_reserved_gb": 26.623046875
}

Inference Evaluation

A short H100 NVL comparison was run against the base model and this adapter using three prompts: Oka frontier path, arithmetic reasoning, and Japanese instruction.

Base average throughput: 5.0904 tok/s
Adapter average throughput: 4.9798 tok/s
Adapter overhead vs base: 2.17%

The adapter solved the arithmetic prompt correctly, but the Japanese instruction prompt showed repetition/hallucinated training-token claims. This checkpoint should be treated as a training-progress artifact, not a production-quality model.

The raw evaluation artifact is included as inference_eval_20260509.json.

Usage

python

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

base_model = "google/gemma-3-27b-it"
adapter = "com-junkawasaki/oka-gemma3-27b-hf-quality-lora"

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

Limitations

Early 100-step checkpoint only.
Not safety tuned.
Not validated against full Opus/GPT-5.5 class benchmark suites.
Japanese output quality regressed in the short eval and needs additional teacher-distillation/preference work.
Does not include merged base weights.