Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled

This repository contains the merged 16-bit supervised fine-tuning checkpoint and GGUF exports for a dense Qwen 3.5 reasoning model based on unsloth/Qwen3.5-27B. It is adapted from the official Unsloth A100 (80GB) notebook and trained on a cleaned subset of nohurry/Opus-4.6-Reasoning-3000x-filtered.

Repository contents

• Merged 16-bit model weights for Transformers / vLLM-style deployment
• GGUF exports for llama.cpp-compatible runtimes
• GGUF quantizations uploaded by the notebook: q4_k_m, q8_0, q5_k_m

Training data

Raw dataset statistics:

• Total rows in train: 2,326
• Columns: id, problem, thinking, solution, difficulty, category, timestamp, hash

Cleaning and formatting applied in the notebook:

• Removed 18 rows with an empty problem, thinking, or solution
• Removed 92 meta / incomplete-prompt responses
• Removed 33 duplicate id rows
• Final training set size: 2,183 conversations
• Category mix after cleaning: 2,052 math, 131 code

Each training example is converted to Qwen chat format with the assistant target built as:

<think>
{thinking}
</think>

{solution}

Loss is applied only to assistant tokens via train_on_responses_only.

Training procedure

• Base model: unsloth/Qwen3.5-27B
• Reference notebook: Unsloth Qwen_3_5_27B_A100(80GB).ipynb
• Frameworks: Unsloth, TRL, Transformers, PEFT
• Task: supervised fine-tuning for long-form reasoning / answer generation
• max_seq_length: 4096
• LoRA rank: 8
• LoRA target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, out_proj
• per_device_train_batch_size: 4
• gradient_accumulation_steps: 2
• num_train_epochs: 1
• learning_rate: 2e-4
• warmup_steps: 20
• optim: adamw_8bit
• weight_decay: 0.001
• lr_scheduler_type: linear
• Seed: 3407

Intended use

This repository is intended for research and experimentation on reasoning-style text generation, especially mixed math and code-oriented prompts that benefit from multi-step intermediate reasoning. The merged checkpoint is suitable for Transformers / vLLM-style serving, and the GGUF files are intended for llama.cpp-compatible runtimes.

Limitations

• Quantized GGUF variants may behave differently from the merged 16-bit checkpoint.
• The model was trained on explicit reasoning traces and may emit visible <think> sections or long intermediate reasoning.
• No formal evaluation or benchmark scores are included in this release.

Usage

Transformers / merged checkpoint:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo_id = "Ayodele01/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else torch.float16,
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Solve: If 3x + 5 = 20, what is x?"}
]

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

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    do_sample=False,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))

GGUF / llama.cpp:

./llama-cli -m Qwen3.5-27B-opus46-reasoning.Q4_K_M.gguf -p "Solve: If 3x + 5 = 20, what is x?" -n 512

Provenance

• Base model: unsloth/Qwen3.5-27B
• Dataset: nohurry/Opus-4.6-Reasoning-3000x-filtered
• Reference notebook source: unslothai/notebooks