LYU2609

InterWebCoder SFT Adapter - step225

This repository contains the LoRA adapters for the local InterWebCoder reconstruction training runs. The latest adapter is:

sft/interwebcoder_adapter_step225

It is an adapter, not a complete base model. Load it on the same Qwen3.5-9B base model used for training, or override the base model path recorded in sft/interwebcoder_adapter_step225/adapter_config.json.

Training

• Base model: local /data/models/Qwen3.5-9B
• Data: 90 verified functional_pass InterWebCoder cases (not included here)
• Prompt source: html_context_test_plan.json and the reference HTML
• HTML interaction and assertion convention: data-testid
• LoRA rank 16, alpha 32, dropout 0
• 5 epochs, 225 optimizer steps
• bfloat16, FSDP2, sequence parallel size 4, maximum sequence length 16,384
• Final-step loss 0.4668, final token accuracy 0.8472
• Adapter tensors: 496; trainable parameters: 43,278,336

Download

hf download DesonDai/LYU2609 \
  --include 'sft/interwebcoder_adapter_step225/*' \
  --local-dir /path/to/adapter

Use the downloaded sft/interwebcoder_adapter_step225 directory as SFT_ADAPTER in the project GRPO launcher. The dataset, screenshots, and base model are intentionally not uploaded to this model repository.