ermiaazarkhalili/Qwen3.5-0.8B-SFT-Claude-Opus-Reasoning-Unsloth

Qwen3.5-0.8B-SFT-Unsloth — Fine-tuned on Claude-Opus-Reasoning

Fine-tuned adapter-merged checkpoint of Qwen3.5-0.8B on the ermiaazarkhalili/Claude-Opus-4.7-Reasoning dataset, produced via Unsloth + Hugging Face TRL's SFTTrainer.

Field	Value
Base model	unsloth/Qwen3.5-0.8B
Architecture	Qwen3ForCausalLM
Parameters	0.8B
Precision	bfloat16 (merged 16-bit)
Fine-tuning method	QLoRA (4-bit base, LoRA r=16, α=16)
Dataset	ermiaazarkhalili/Claude-Opus-4.7-Reasoning (distillation corpus examples)
Training	N=1 full epoch (N=1 epoch steps, effective batch=8)
Learning rate	2e-4 (linear decay, warmup 5 steps)
Unsloth version	2026.4.6
Trained on	DRAC Fir cluster, NVIDIA H100 80GB HBM3 MIG 3g.40gb

Usage

Python (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ermiaazarkhalili/Qwen3.5-0.8B-SFT-Claude-Opus-Reasoning-Unsloth",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("ermiaazarkhalili/Qwen3.5-0.8B-SFT-Claude-Opus-Reasoning-Unsloth", trust_remote_code=True)

messages = [{"role": "user", "content": "Explain step-by-step: if a train travels 60 mph for 2.5 hours, how far does it go?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text=text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Unsloth (2× faster inference)

from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    "ermiaazarkhalili/Qwen3.5-0.8B-SFT-Claude-Opus-Reasoning-Unsloth",
    max_seq_length=2048,
    load_in_4bit=False,
)
FastLanguageModel.for_inference(model)

GGUF (llama.cpp / Ollama)

Quantized GGUF versions are available at ermiaazarkhalili/Qwen3.5-0.8B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF:

# llama-cli
llama-cli -hf ermiaazarkhalili/Qwen3.5-0.8B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF --jinja -p "Explain step-by-step: if a train travels 60 mph for 2.5 hours, how far does it go?" -n 256

# Ollama
ollama run hf.co/ermiaazarkhalili/Qwen3.5-0.8B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF:Q4_K_M

Training details

Reasoning SFT fine-tuning on distillation corpus examples from ermiaazarkhalili/Claude-Opus-4.7-Reasoning.

• SFTTrainer (trl >= 0.14) via Unsloth's FastLanguageModel
• LoRA config: r=16, α=16, dropout=0, targeting q_proj/k_proj/v_proj/o_proj/gate_proj/up_proj/down_proj
• Effective batch size: 8 (per_device=2 × grad_accum=4)
• Max sequence length: 2048
• Optimizer: adamw_8bit with linear LR scheduler
• Seed: 3407

Intended use

For research and non-commercial experimentation only. Outputs should be independently verified before any downstream use.

Limitations

• Trained on a single epoch (~N=1 epoch optimizer steps); further training may yield additional gains.
• Fine-tuned from Qwen3.5-0.8B, inherits its limitations and biases.
• Evaluated only on training-data perplexity; no external benchmarks run on this checkpoint.
• Distilled reasoning traces reflect patterns from Claude Opus 4.7 and may not generalize to domains outside the distillation corpus.

Citation

@misc{ qwen35_08b_sft_claude_opus_2026 ,
  author = {Ermia Azarkhalili},
  title = { Qwen3.5-0.8B-SFT-Unsloth — Reasoning SFT fine-tune of Qwen3.5-0.8B },
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ermiaazarkhalili/Qwen3.5-0.8B-SFT-Claude-Opus-Reasoning-Unsloth}}
}

---

This qwen3 model was trained 2× faster with Unsloth and Hugging Face's TRL library.