RessAI/Onner-O1-Qwen35-4B-Reasoning

Onner-O1 Qwen3.5 4B Reasoning

RessAI/Onner-O1-Qwen35-4B-Reasoning is a merged Transformers checkpoint derived from Qwen/Qwen3.5-4B.

This model was post-trained for reasoning-style interaction, Indonesian/English instruction following, tool-call formatting, and long-context experimentation. The uploaded Transformers weights are merged; no external LoRA adapter is required.

Model Details

• Base model: Qwen/Qwen3.5-4B
• Released model id: RessAI/Onner-O1-Qwen35-4B-Reasoning
• Format: merged Hugging Face Transformers checkpoint
• Training type: BF16 LoRA supervised fine-tuning + supervised repair
• Not trained with: DPO or GRPO
• Primary languages: Indonesian and English
• Intended use: research, local inference, API serving experiments, reasoning/tool-use prototyping

Important Limitations

This is an experimental post-trained model. It may still:

• repeat words or phrases under poor decoding settings;
• enter self-doubt or thought loops on hard reasoning prompts;
• answer in the wrong language if the system prompt is written only in English;
• hallucinate unsupported facts;
• fail exact-answer benchmark formatting;
• degrade on very long contexts without careful runtime configuration.

Use conservative decoding and validate outputs for important tasks.

Recommended Decoding

For reasoning/math:

temperature = 0.2
top_p = 0.8
top_k = 20
repetition_penalty = 1.08
max_new_tokens = 2048

For direct/no-thinking answers:

temperature = 0.0
top_p = 1.0
repetition_penalty = 1.05
max_new_tokens = 512

Recommended system instruction:

Ikuti bahasa pengguna.
Jika pengguna bertanya dalam bahasa Indonesia, jawab dalam bahasa Indonesia.
Jika pengguna bertanya dalam bahasa Inggris, jawab dalam bahasa Inggris.

Bernalar secara ringkas.
Jangan ragu-ragu berulang.
Jika menemukan kesalahan, koreksi sekali lalu lanjut.
Jangan mengulang frasa seperti "wait", "maybe", atau "let me reconsider".
Akhiri dengan satu jawaban final yang jelas.

Transformers Usage

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "RessAI/Onner-O1-Qwen35-4B-Reasoning"

processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {
        "role": "system",
        "content": (
            "Ikuti bahasa pengguna. Bernalar secara ringkas. "
            "Jangan mengulang frasa atau ragu-ragu berulang."
        ),
    },
    {"role": "user", "content": "Jawab hanya dengan angka: 17 dikali 23 berapa?"},
]

inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    enable_thinking=False,
).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=128,
    do_sample=False,
    repetition_penalty=1.05,
)

print(processor.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Expected answer:

391

Thinking Mode

The chat template supports enable_thinking=True or enable_thinking=False.

Use enable_thinking=True for math, science, and multi-step reasoning. Use enable_thinking=False for short factual answers, extraction, and low-latency responses.

GGUF

A Q4_K_M GGUF build is included under the gguf/ folder when available:

gguf/qwen35-4b-reasoning-1m-Q4_K_M.gguf

Recommended local settings:

• context: 2K, 4K, or 8K for normal consumer devices;
• avoid 262K/1M context on laptops/phones;
• use a repetition penalty around 1.05 to 1.10.

Evaluation Notes

Raw local evaluation artifacts were generated for:

• AIME 2026
• HMMT Nov 2025 / Feb 2026
• SuperGPQA
• SciBench
• MBPP
• LongBench v2
• IMOAnswerBench
• CritPt

However, these raw numbers should not be treated as leaderboard-comparable scores without rerunning a corrected evaluation harness. Known issues in the previous evaluation include:

• HMMT outputs were often truncated by low max_new_tokens;
• SciBench numeric grading was too strict;
• MBPP prompts did not always preserve required function names;
• some exact-answer tasks failed due to final-answer formatting rather than model capability alone.

Training Summary

The local training ledger recorded:

• Phase 1 SFT/LoRA runtime: about 1284 seconds
• Phase 1 train loss: about 0.7563
• Phase 2 SFT/LoRA runtime: about 282 seconds
• Phase 2 train loss: about 1.7346
• Repair SFT steps: 30
• Repair train loss: about 0.3085

Safety and Use

This model is provided for experimentation. Do not rely on it for legal, medical, financial, safety-critical, or high-stakes decisions without independent verification.

License

This repository follows the license terms of the base model where applicable. See the base model Qwen/Qwen3.5-4B for upstream license and usage requirements.