CoolFace
Modelpublic

ludolara/fp8-sft-Llama3.3-70B

sourceHugging Faceupdated 4mo agoView on Hugging Face
0likes8downloads
Model Card

fp8-sft-Llama3.3-70B

LoRA SFT adapter for generating 8-room RPLAN floor plans from structured room and adjacency specifications.

Model Details

  • —Base model: meta-llama/Llama-3.3-70B-Instruct
  • —Local training base path: models/Llama-3.3-70B-Instruct
  • —Adapter type: LoRA / PEFT
  • —LoRA rank: 64
  • —LoRA alpha: 128
  • —Training stage: supervised fine-tuning (SFT)
  • —Task: RPLAN 8-room floor-plan generation
  • —Local source: output/final/rplan8_2_70B_r64_a128_all

Reported SFT Metrics

MetricValue
Room Area error0.08 +/- 0.05
Room ID recall1.00 +/- 0.00
Overlap0.37 +/- 0.48
Percent overlap0.01 +/- 0.03
Compatibility0.41 +/- 0.73
Diversity6.44 +/- 0.00

Intended Use

Use this adapter with the Llama-3.3-70B-Instruct base model to generate 8-room residential floor plans in the JSON format used by the paper. It is intended for research and evaluation on RPLAN-style floor-plan generation.

Limitations

The model is specialized to the 8-room RPLAN task and may not generalize to other datasets, room-count regimes, architectural standards, or safety-critical design workflows.