--- library_name: transformers license: other license_name: lfm1.0 base_model: LiquidAI/LFM2-2.6B base_model_relation: finetune pipeline_tag: text-generation language: - en tags: - chemistry - retrosynthesis - chemical-synthesis - single-step-retrosynthesis - synthetic-dataset - virtual-synthesis ---
# C3LM **Chemistry Constraint–Consistent Language Model** [![Paper](https://img.shields.io/badge/Paper-arXiv%3A2602.03554-B31B1B?style=flat-square&logo=arxiv&logoColor=white)](https://arxiv.org/abs/2602.03554) [![ChemCensor](https://img.shields.io/badge/github-ChemCensor-blue?logo=github)](https://github.com/insilicomedicine/ChemCensor)
CREED overview
--- ## Overview C3LM is a chemistry-focused language model for Single-Step Retrosynthesis task. This checkpoint was initialized from **LFM2-2.6B** and trained with **supervised fine-tuning (SFT)** on a combined **CREED + USPTO-full** dataset. --- ## Load model ```python from transformers import AutoModelForCausalLM, AutoTokenizer # Load model and tokenizer model_id = "insilicomedicine/C3LM" model = AutoModelForCausalLM.from_pretrained( model_id, device_map="auto", torch_dtype="bfloat16", # attn_implementation="flash_attention_2" <- uncomment on compatible GPU ) tokenizer = AutoTokenizer.from_pretrained(model_id) ``` --- ## Usage example ```python import re ATOM_TOKENS = ["Ag","Al","As","Au","Bi","Br","Ca","Cl","Cr","Cu","Fe","Ga","Gd","Ge","Hg","Li","Mg","Mo","Na","Pt","Ru","Sb","Se","Si","Sn","Sc","B","C","F","H","I","K","M","N","O","P","S","V","W","Z","c","e","n","o","p","s"] SMI_REGEX_PATTERN = ( r"(\[|\]|\(|\)|\.|=|#|-|\+|\\|\/|:|~|@|\?|>>?|\*|\$|\%[0-9]{2}|[0-9]|" + "|".join(ATOM_TOKENS) + ")" ) smi_regex = re.compile(SMI_REGEX_PATTERN) def isolate_sequence(s): return "".join(["" for tok in smi_regex.findall(s)]) def deisolate_sequence(s): s = re.sub("[<\s]?sm_[^>]+>", lambda matchobj: matchobj.group(0)[4:-1], s).strip() return s # Prepare prompts product_smiles = "N1C(=O)Cc2ccc(Nc3ccc(N[C@@H]4CN(C(c5cc(NCCO)ccc5)=O)C[C@@H]4c4scnc4)cc3)cc21" prompt = f"What are the possible reactants that could have formed the following product {isolate_sequence(product_smiles)}? Wrap an answer in tags/think" input_ids = tokenizer.apply_chat_template( [ {"role": "system", "content": "You are a helpful assistant in chemistry and biology."}, {"role": "user", "content": prompt} ], add_generation_prompt=True, return_tensors="pt", tokenize=True, )['input_ids'].to(model.device) # Generate answer output = model.generate( input_ids, do_sample=True, temperature=0.3, min_p=0.15, repetition_penalty=1.05, max_new_tokens=512, ) for out in output: out = tokenizer.decode(out, skip_special_tokens=False) answer_part = out.split('')[1].split('')[1].split('')[0] answer_part = deisolate_sequence(answer_part) print('Reactants: ', answer_part) ``` --- ## Citation ``` @misc{zagribelnyy2026singleanswerenoughrethinking, title={When Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs}, author={Bogdan Zagribelnyy and Ivan Ilin and Maksim Kuznetsov and Nikita Bondarev and Roman Schutski and Thomas MacDougall and Rim Shayakhmetov and Zulfat Miftakhutdinov and Mikolaj Mizera and Vladimir Aladinskiy and Alex Aliper and Alex Zhavoronkov}, year={2026}, eprint={2602.03554}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2602.03554}, } ```