retro-agent-qwen25-7b-merged-fp16

Standalone fp16 retrosynthesis assistant model for serving. The model is based on Qwen/Qwen2.5-7B-Instruct with these LoRA adapters merged into the base weights:

• retro-reactants-qwen25-7b
• retro-conditions-qwen25-7b

Because the LoRA weights are merged, deployment does not need PEFT adapter loading. Load and serve this repository as a normal causal language model.

Intended Use

This model is intended for structured retrosynthesis assistance:

• predict likely reactants for a target product SMILES
• predict plausible reaction conditions
• return machine-readable JSON for downstream validation and rendering

Use RDKit or another chemistry toolkit to validate generated SMILES before acting on model output.

Prompt Templates

Use explicit task labels and request JSON only. Low temperature is recommended for repeatable API calls.

Reactant Prediction

System: You are a retrosynthesis assistant. Return valid JSON only.
User: Task: predict_reactants
Product SMILES: CC(=O)Oc1ccccc1C(=O)O
Return JSON with keys: reactants_smiles, confidence, rationale.

Condition Prediction

System: You are a reaction condition assistant. Return valid JSON only.
User: Task: predict_conditions
Product SMILES: CC(=O)Oc1ccccc1C(=O)O
Reactants SMILES: CC(=O)O.Oc1ccccc1C(=O)O
Return JSON with keys: solvent, reagents, catalyst, temperature_celsius, time, confidence, rationale.

Combined Planning Prompt

System: You are a retrosynthesis assistant. Return valid JSON only.
User: Task: retrosynthesis_step
Target product SMILES: CC(=O)Oc1ccccc1C(=O)O
Return one JSON object with keys: reactants_smiles, conditions, confidence, rationale.

Python Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "<your-namespace>/retro-agent-qwen25-7b-merged-fp16"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "You are a retrosynthesis assistant. Return valid JSON only."},
    {"role": "user", "content": "Task: retrosynthesis_step\nTarget product SMILES: CC(=O)Oc1ccccc1C(=O)O\nReturn one JSON object with keys: reactants_smiles, conditions, confidence, rationale."},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.1,
        do_sample=False,
    )

new_tokens = output_ids[0][inputs.input_ids.shape[-1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

REST API Examples

Hugging Face Router / OpenAI-compatible Chat Completions

This works only if the model is available through Hugging Face Inference Providers. For custom merged models, a dedicated endpoint is often more reliable.

export HF_TOKEN=hf_...

curl https://router.huggingface.co/v1/chat/completions \
  -H "Authorization: Bearer $HF_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<your-namespace>/retro-agent-qwen25-7b-merged-fp16",
    "messages": [
      {"role": "system", "content": "You are a retrosynthesis assistant. Return valid JSON only."},
      {"role": "user", "content": "Task: retrosynthesis_step\nTarget product SMILES: CC(=O)Oc1ccccc1C(=O)O"}
    ],
    "temperature": 0.1,
    "max_tokens": 512,
    "stream": false
  }'

Dedicated Hugging Face Inference Endpoint

Create a dedicated endpoint from this model repository, then call the endpoint URL directly.

export HF_TOKEN=hf_...
export ENDPOINT_URL=https://<your-endpoint>.<region>.<provider>.endpoints.huggingface.cloud

curl "$ENDPOINT_URL" \
  -X POST \
  -H "Authorization: Bearer $HF_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "inputs": "Task: retrosynthesis_step\nTarget product SMILES: CC(=O)Oc1ccccc1C(=O)O\nReturn one JSON object with keys: reactants_smiles, conditions, confidence, rationale.",
    "parameters": {
      "max_new_tokens": 512,
      "temperature": 0.1,
      "return_full_text": false
    }
  }'

Notes

This model merges adapters trained for related but distinct tasks. Evaluate both reactant prediction and condition prediction after merge, because independently trained LoRA adapters can interfere with each other.