SM_heat_exchanger_v1

LoRA adapter for Qwen/Qwen2.5-7B-Instruct that designs shell-and-tube and concentric-tube heat exchangers as JSON.

It was trained with GRPO against the heat_exchanger environment of sm-heat-exchanger v0.8.0 (on sm-core v0.8.0): a deterministic physics referee that checks each design (schema, design rules, simulation) and returns a score in [0, 1].

This is the first Suni Muhendis model. Its results on design families it never saw during training will be published on SM-Bench, next to other models and the base model.

Usage

  • Prompt: sm_heat_exchanger.prompts.build_heat_exchanger_prompt(task), sent as a single user chat message.
  • Decoding: greedy.
  • Score the output with sm_core.make_env("heat_exchanger").evaluate(...).
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct", device_map="auto")
model = PeftModel.from_pretrained(base, "SM-AI/SM_heat_exchanger_v1")
tokenizer = AutoTokenizer.from_pretrained("SM-AI/SM_heat_exchanger_v1")

Limitations

  • Heat exchangers only, water on both sides, within the ranges of the sm-heat-exchanger referee. Designs are scored by a simplified model, not certified for construction.

License

Apache-2.0, as the base model.

Downloads last month
17
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SM-AI/SM_heat_exchanger_v1

Base model

Qwen/Qwen2.5-7B
Adapter
(2,934)
this model