qwen_llm_scs — Qwen3.5-VL 27B continued-pretrained on spinal cord stimulation literature

Full multimodal (text + vision) model. A LoRA adapter trained on spinal cord stimulation (SCS) papers has been merged into the base Qwen/Qwen3.6-27B weights. The vision tower is unchanged from the base; only the language layers were adapted.

Model details

• Architecture: Qwen3_5ForConditionalGeneration (model_type: qwen3_5)
• Base model: Qwen/Qwen3.6-27B (full VLM)
• Adaptation: LoRA r=16, alpha=32, dropout 0.05, continued pre-training
• LoRA targets: q_proj, k_proj, v_proj, o_proj, out_proj, gate_proj, up_proj, down_proj (language layers only — vision tower not touched)
• Precision: BF16 (base FP8 dequantized at load time, then LoRA merged)
• Size: ~51 GB, 12 safetensors shards
• Domain: spinal cord stimulation clinical and engineering literature

Usage

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

repo = "achuthc1298/qwen_llm_scs"

processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    repo,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
    attn_implementation="sdpa",
)
model.eval()

# Text-only
messages = [{"role": "user", "content": [{"type": "text", "text": "Summarize the principle of high-frequency SCS."}]}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=400, do_sample=False)
print(processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

# Vision (figure from a paper)
from PIL import Image
img = Image.open("figure.png").convert("RGB")
messages = [{"role": "user", "content": [
    {"type": "image", "image": img},
    {"type": "text", "text": "Describe this figure."},
]}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=400, do_sample=False)
print(processor.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Hardware

Tested on 2× RTX A6000 (48 GB each) with device_map="auto" and per-GPU memory limits of 44 GiB. Total VRAM at inference ≈ 57 GB in BF16.

Notes

• The vision tower (model.visual.*) is identical to the base model — only the language layers received SCS-domain LoRA updates.
• Loading uses the native qwen3_5 integration in modern transformers; no custom remote code is bundled.
• The chat template is the standard Qwen3-VL template.

License

Inherits the Qwen license of the base model.