Instructions to use jeffchanpm/Qwen3.5-27B-SGR-LCL with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use jeffchanpm/Qwen3.5-27B-SGR-LCL with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="jeffchanpm/Qwen3.5-27B-SGR-LCL")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("jeffchanpm/Qwen3.5-27B-SGR-LCL")
model = AutoModelForMultimodalLM.from_pretrained("jeffchanpm/Qwen3.5-27B-SGR-LCL", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use jeffchanpm/Qwen3.5-27B-SGR-LCL with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "jeffchanpm/Qwen3.5-27B-SGR-LCL"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jeffchanpm/Qwen3.5-27B-SGR-LCL",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/jeffchanpm/Qwen3.5-27B-SGR-LCL

SGLang

How to use jeffchanpm/Qwen3.5-27B-SGR-LCL with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "jeffchanpm/Qwen3.5-27B-SGR-LCL" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jeffchanpm/Qwen3.5-27B-SGR-LCL",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "jeffchanpm/Qwen3.5-27B-SGR-LCL" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "jeffchanpm/Qwen3.5-27B-SGR-LCL",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use jeffchanpm/Qwen3.5-27B-SGR-LCL with Docker Model Runner:
```
docker model run hf.co/jeffchanpm/Qwen3.5-27B-SGR-LCL
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Qwen3.5-27B-SGR-LCL

Fine-tuned from Qwen3.5-27B using Self-Graph Reasoning (SGR) + Logical Curriculum Learning (LCL) — a 4-phase progressive training approach based on LoCM (Logical Complexity Metric).

Highlights

Thinking mode: +6.5% overall improvement over Qwen3.5-27B baseline (78.0% → 84.5%)
ProverQA Hard: +12.0% in thinking mode (62.0% → 74.0%)
Near-zero skip-think degradation: only -1.2% (improved from -5.7% of naive LoRA)

Method

Self-Graph Reasoning (SGR)

SGR is a graph-structured reasoning framework that enables LLMs to explicitly represent their reasoning processes as structured graphs prior to producing final answers. See arXiv:2601.03597.

Logical Curriculum Learning (LCL)

Standard SGR training data has a mean LoCM of 23.33 with zero samples below LoCM 6, causing the model to skip foundational logic. LCL addresses this with 4-phase progressive training:

Phase	LoCM Range	Data Source	Samples
1	0-6	ProntoQA + ProofWriter	360
2	6-10	ProntoQA + ProofWriter + SFT	724 (cumulative)
3	10-15	Original SFT	996 (cumulative)
4	15-30	Original SFT	9,536 (cumulative)

Each phase warm-starts from the previous phase's best LoRA checkpoint. LoCM (Logical Complexity Metric) is from arXiv:2601.02902.

Training Details

Hardware: NVIDIA DGX Spark (Grace Blackwell GB10, 128 GB unified memory)
Framework: Unsloth + PEFT (BF16, LoRA r=16, alpha=32)
LoRA targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (256 pairs across 64 layers)
Optimizer: AdamW 8-bit, lr=2e-5, cosine schedule
Batch: effective batch size 8 (bs=1 × grad_accum=8)
Total training: ~24 hours across 4 phases

Merge

LoRA merged into base weights via manual matrix multiplication (W_merged = W_base + B @ A * scaling), bypassing PEFT merge_and_unload() which has a key mapping issue with Qwen3.5's language_model prefix.

Results (LPT Benchmark)

Thinking Mode (200 questions, 50 per dataset)

Dataset	Qwen3.5-27B	SGR-LCL	Delta
ProverQA Easy	92.0%	96.0%	+4.0%
ProverQA Medium	76.0%	82.0%	+6.0%
ProverQA Hard	62.0%	74.0%	+12.0%
NSA-LR	82.0%	86.0%	+4.0%
Overall	78.0%	84.5%	+6.5%

Skip-think Mode (3,000 questions)

Dataset	N	Qwen3.5-27B	SGR-LCL	Delta
NSA-LR	1500	76.5%	75.2%	-1.3%
ProverQA Easy	500	85.0%	86.0%	+1.0%
ProverQA Medium	500	74.8%	74.6%	-0.2%
ProverQA Hard	500	68.0%	64.0%	-4.0%
Overall	3000	76.2%	75.0%	-1.2%

Data Contamination

ProverQA and NSA-LR evaluation datasets have zero overlap with training data. Full audit available in the project repository.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "jeffchanpm/Qwen3.5-27B-SGR-LCL",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("jeffchanpm/Qwen3.5-27B-SGR-LCL")

prompt = "If all dogs are animals and all animals are living things, are all dogs living things? Think step by step."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

GGUF Quantized Version

For GGUF quantized versions (Q4_K_M), see jeffchanpm/Qwen3.5-27B-SGR-LCL-GGUF.

Citation

@article{chen2026chains,
  title={From Chains to Graphs: Self-Structured Reasoning for General-Domain LLMs},
  author={Chen, Yingjian and Liu, Haoran and Liu, Yinhong and Tong, Sherry T and Feng, Aosong and Lu, Jinghui and Zhang, Juntao and Iwasawa, Yusuke and Matsuo, Yutaka and Li, Irene},
  journal={arXiv preprint arXiv:2601.03597},
  year={2026}
}

Acknowledgments

Qwen Team for Qwen3.5-27B
Unsloth for efficient fine-tuning on DGX Spark
Logical Phase Transitions for the LoCM complexity metric

Downloads last month: 5

Safetensors

Model size

28B params

Tensor type

BF16

F32

Model tree for jeffchanpm/Qwen3.5-27B-SGR-LCL

Base model

Qwen/Qwen3.5-27B

Adapter

(99)

this model

Quantizations

1 model

Dataset used to train jeffchanpm/Qwen3.5-27B-SGR-LCL

Papers for jeffchanpm/Qwen3.5-27B-SGR-LCL

From Chains to Graphs: Self-Structured Reasoning for General-Domain LLMs

Paper • 2601.03597 • Published Jan 7

Logical Phase Transitions: Understanding Collapse in LLM Logical Reasoning

Paper • 2601.02902 • Published Jan 6 • 11

Evaluation results

Overall Accuracy on LPT Benchmark (Thinking)
self-reported

84.500
ProverQA Hard Accuracy on LPT Benchmark (Thinking)
self-reported

74.000
Overall Accuracy on LPT Benchmark (Skip-think)
self-reported

75.000