UCL-CSSB/PlasmidGPT-GRPO

1import torch2from transformers import AutoTokenizer, AutoModelForCausalLM3 4device = 'cuda' if torch.cuda.is_available() else 'cpu'5print(f"Using device: {device}\n")6 7print("Loading RL-optimized PlasmidGPT-GRPO model...")8model = AutoModelForCausalLM.from_pretrained(9    ".",10    trust_remote_code=True11).to(device)12model.eval()13 14tokenizer = AutoTokenizer.from_pretrained(15    ".",16    trust_remote_code=True17)18 19print("Generating optimized plasmid sequences...\n")20 21start_sequence = 'ATGGCTAGCGAATTCGGCGCGCCT'22print(f"Start sequence: {start_sequence}\n")23 24input_ids = tokenizer.encode(start_sequence, return_tensors='pt').to(device)25 26outputs = model.generate(27    input_ids,28    max_length=400,29    num_return_sequences=3,30    temperature=0.8,31    do_sample=True,32    top_k=50,33    top_p=0.95,34    pad_token_id=tokenizer.pad_token_id,35    eos_token_id=tokenizer.eos_token_id36)37 38print("=" * 80)39for i, output in enumerate(outputs, 1):40    sequence = tokenizer.decode(output, skip_special_tokens=True)41    print(f"\nPlasmid {i}:")42    print(f"  Length: {len(sequence)} bp")43    print(f"  First 100 bp: {sequence[:100]}")44    print(f"  Last 100 bp: {sequence[-100:]}")45print("\n" + "=" * 80)46 47print("\nNote: These sequences are generated by an RL-optimized model trained to:")48print("  ✓ Include proper genetic elements (ori, promoters, CDS, markers)")49print("  ✓ Avoid repeat regions > 50 bp")50print("  ✓ Generate compact, functional plasmids")51print("  ✓ Organize genes in proper cassettes (promoter → CDS → terminator)")52