--- base_model: ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth tags: - gguf - llama.cpp - unsloth - qwen3.5 - sft - distillation - reasoning - quantized license: apache-2.0 language: - en datasets: - ermiaazarkhalili/claude-reasoning-distillation pipeline_tag: text-generation --- # Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth — GGUF quantized GGUF quantizations of [`ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth`](https://huggingface.co/ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth), produced via [Unsloth](https://github.com/unslothai/unsloth) + llama.cpp's conversion scripts. | Field | Value | |---|---| | **Source checkpoint** | [`ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth`](https://huggingface.co/ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth) | | **Base model** | [`unsloth/Qwen3.5-9B`](https://huggingface.co/unsloth/Qwen3.5-9B) | | **Dataset** | [`claude-reasoning-distillation`](https://huggingface.co/datasets/ermiaazarkhalili/claude-reasoning-distillation) (10,477 samples) | | **Architecture** | Qwen3.5 hybrid transformer + state-space (SSM) blocks | | **Training** | 1 full epoch (effective batch=8 via per_device=1 × grad_accum=8) | | **LoRA** | r=64, α=64 (Jackrong recipe), dropout=0 | | **Final training loss** | 0.795 (job 37216727, runtime 2h 32m on H100 MIG 3g.40gb, peak VRAM 20.25 GB) | | **Conversion** | Unsloth `save_pretrained` → llama.cpp `convert_hf_to_gguf.py` → `llama-quantize` | | **Quantization tool** | llama.cpp `llama-quantize` (cached toolchain) | ## Available quantizations | File | Bits | Size | Notes | |---|---|---|---| | `qwen3.5-9b-sft-claude-opus-reasoning-unsloth.q2_k.gguf` | 2-bit | 3.83 GB | Smallest; aggressive quality loss | | `qwen3.5-9b-sft-claude-opus-reasoning-unsloth.q3_k_m.gguf` | 3-bit | 4.62 GB | Small; noticeable quality loss | | `qwen3.5-9b-sft-claude-opus-reasoning-unsloth.q4_k_m.gguf` | 4-bit | 5.63 GB | **Recommended** — best size/quality balance | | `qwen3.5-9b-sft-claude-opus-reasoning-unsloth.q5_k_m.gguf` | 5-bit | 6.47 GB | Near-full quality | | `qwen3.5-9b-sft-claude-opus-reasoning-unsloth.q6_k.gguf` | 6-bit | 7.36 GB | Very close to Q8_0 at smaller size | | `qwen3.5-9b-sft-claude-opus-reasoning-unsloth.q8_0.gguf` | 8-bit | 9.53 GB | Largest; closest to bf16 source | **Recommended default:** `Q4_K_M` (4-bit, K-quant medium). For maximum fidelity, use `Q8_0`. ## Usage ### llama.cpp ```bash # Reasoning-style chat llama-cli -hf ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF --jinja -p "Solve step by step: What is the sum of the first 10 prime numbers?" -n 512 # Interactive llama-cli -hf ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF --jinja -cnv ``` ### Ollama ```bash ollama run hf.co/ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF:Q4_K_M ``` ### llama-cpp-python ```python from llama_cpp import Llama llm = Llama.from_pretrained( repo_id="ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF", filename="*q4_k_m.gguf", n_ctx=2048, ) out = llm.create_chat_completion( messages=[{"role": "user", "content": "Solve step by step: What is the sum of the first 10 prime numbers?"}], max_tokens=512, ) print(out["choices"][0]["message"]["content"]) ``` ## Intended use For research and non-commercial experimentation in chain-of-thought reasoning tasks. Distilled from Claude Opus reasoning traces (` ` blocks) to reproduce step-by-step problem solving. Outputs should be independently verified before any downstream use. ## Limitations - GGUF quantizations have unavoidable quality loss relative to the source bfloat16 checkpoint. Use `Q5_K_M`, `Q6_K`, or `Q8_0` for best fidelity. - Inherits all limitations of the source merged checkpoint ([`ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth`](https://huggingface.co/ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth)). - Trained on Claude reasoning traces — inherits any biases or failure modes present in those traces. - The Qwen3.5 architecture includes SSM (state-space model) blocks; some llama.cpp builds prior to SSM support may not run these GGUFs correctly. Use a recent llama.cpp build. ## Citation ```bibtex @misc{ qwen35_9b_sft_claude_unsloth_2026_gguf , author = {Ermia Azarkhalili}, title = { Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth — GGUF quantized }, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/ermiaazarkhalili/Qwen3.5-9B-SFT-Claude-Opus-Reasoning-Unsloth-GGUF}} } ``` --- This qwen3.5 model was trained 2× faster with [Unsloth](https://github.com/unslothai/unsloth) and Hugging Face's TRL library. [ ](https://github.com/unslothai/unsloth)
