PrimeMind-0.8B
Compressed reasoning model based on Qwen/Qwen3.5-0.8B, fine-tuned on caveman-style thinking datasets.
What it does
Teaches the model to think in short, concise, information-packed style using <think> tags, then output clean answers.
Training
- Base model: Qwen/Qwen3.5-0.8B
- Method: LoRA SFT (4-bit NF4 quantization)
- Datasets:
catsaresupercool/synthetic-caveman-thinking(600 samples)nibauman/objectnav-sft-claude-caveman(600 samples, multimodal)
- Total samples: 1,200
- Epochs: 1
- Loss: 4.81 → 1.56
- Training time: ~38 minutes on RTX 4060 Ti (16GB)
LoRA Config
- Rank: 64
- Alpha: 128
- Dropout: 0.05
- Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Usage
from transformers import AutoProcessor, AutoModelForMultimodalLM
model = AutoModelForMultimodalLM.from_pretrained(
"CrowdMind/PrimeMind-0.8B",
torch_dtype="bfloat16",
device_map="auto",
trust_remote_code=True,
)
processor = AutoProcessor.from_pretrained("CrowdMind/PrimeMind-0.8B", trust_remote_code=True)
messages = [{"role": "user", "content": "Explain quantum computing in one sentence."}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=256)
response = processor.tokenizer.decode(output[0][inputs["input_ids"].shape[1]:])
print(response)