FoamGPT adapters

LoRA adapters that let one open-weight model serve every LLM call of an end-to-end OpenFOAM agent: planning the case, writing the configuration files, diagnosing solver errors, and repairing the case. They accompany the paper FoamGPT: Fine-Tuning Large Language Models for Agentic Automation of CFD Simulations with OpenFOAM (anonymous submission). Code, data, and evaluation scripts are in the companion code repository.

Adapters

Folder	Base model	Training records	In the paper
nemotron-3-nano-30b/shared/	nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16	2,500 (all nine call sites)	FoamGPT-Nemotron (Table 1)
gemma4-26b-a4b/shared/	unsloth/gemma-4-26b-a4b-it	2,499 (all nine call sites)	FoamGPT-Gemma4 (Table 1)
nemotron-3-nano-30b/split/stage-specific/{plan,write,review}/	Nemotron	the records of each group	stage-specific split C (Appendix B)
nemotron-3-nano-30b/split/random-R{1,2,3}/G{1,2,3}/	Nemotron	the records of each group	random splits R1-R3 (Appendix B)
nemotron-3-nano-30b/scaling/{L025,L050,L075,L100}/	Nemotron	560 / 1,314 / 1,874 / 2,500	data-scaling curve (Figure 4)

The shared adapters are the systems reported as FoamGPT. Records are teacher completions from variants 1-7 of the FoamBench Basic families; records longer than 16,384 tokens after chat templating are dropped (147 error_analysis records with the Nemotron tokenizer, 148 with the Gemma4 tokenizer). The manifests of trained records are in the code repository.

Which split adapter serves which call site

Call site	stage-specific	R1	R2	R3
parse_case_info	plan	G1	G1	G3
build_advice	plan	G1	G2	G1
decompose_subtasks	plan	G3	G2	G1
generate_file	write	G2	G2	G2
allrun_commands	write	G2	G1	G3
allrun_script	write	G1	G3	G2
error_analysis	review	G3	G1	G2
rewrite_plan	review	G2	G2	G2
rewrite_files	write	G2	G3	G1

Training

Supervised fine-tuning on the teacher's completions (loss on completion tokens only), 3 epochs, learning rate 2e-4 with a cosine schedule (0.1 minimum-rate floor), warmup ratio 0.05, effective batch 16, weight decay 0.01, seed 3407, maximum length 16,384 tokens, thinking disabled.

• Nemotron adapters: LoRA r = 16, alpha = 32, dropout 0, on q/k/v/o_proj, gate/up/down_proj, and the Mamba in/out_proj. TRL 0.24.0, Transformers 4.56.2, PEFT 0.18.1, PyTorch 2.7.1.
• Gemma4 adapter: LoRA r = 16, alpha = 16, dropout 0, on the language-tower attention and feed-forward linears, trained through Unsloth. TRL 0.24.0, Transformers 5.5.0, PEFT 0.19.1, PyTorch 2.10.0.

Use

Each folder holds adapter_model.safetensors, adapter_config.json, and the tokenizer files and chat_template.jinja the adapter was trained and served with. Example with vLLM:

vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 \
  --enable-lora --max-lora-rank 16 --lora-modules foamgpt=./nemotron-3-nano-30b/shared \
  --chat-template ./nemotron-3-nano-30b/shared/chat_template.jinja \
  --tensor-parallel-size 4 --max-model-len 100000

The paper's runs use temperature 0.6, a 16,384-token output cap, and the agent harness from the code repository. For Gemma4 the adapter was merged into the bfloat16 base model for serving (PEFT merge_and_unload).

License and terms

See LICENSE.md. The Nemotron adapters are distributed under the NVIDIA Nemotron Open Model License and the Gemma4 adapter under Apache-2.0, following their base models. The training data are outputs of Claude Opus 4.6 obtained through Amazon Bedrock and are subject to Anthropic's terms of service and usage policy. The adapters are released as research artifacts for research use.