You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Configuration Parsing Warning:In adapter_config.json: "peft.task_type" must be a string

CodeAnything Qwen3.5-9B SFT 1.835M

LoRA and multimodal-merger checkpoint trained from Qwen3.5-9B Base on the CodeAnything 1.835M SFT dataset. This is the Base-init model used for the final 800-sample, 16-domain evaluation.

Evaluation

Chart	Table	Flow	Math	Web	SVG	Beamer	MARP	Bio	Chem	ChSmiles	Circuit	ABC	CAD	Mesh	Formula	Overall
2.60	3.90	3.96	2.98	2.62	2.92	3.10	2.84	2.20	3.44	4.50	2.70	4.72	2.20	2.68	4.98	3.27

Scores use 800 samples (50 per domain) and the GPT-5.5 detailed visual judge.

Files

adapter_model.safetensors: LoRA weights.
non_lora_state_dict.bin: trained non-LoRA multimodal merger weights.
adapter_config.json: PEFT configuration.
tokenizer, processor, chat template, and model configuration files.

Optimizer, scheduler, RNG, and other training-resume state are intentionally excluded.

Loading

Download this repository and load Qwen3.5-9B Base, then attach the adapter with PEFT. Apply non_lora_state_dict.bin to the base model with strict=False before inference. The companion GitHub release contains download and loading helpers.

Dataset and code

Dataset: Ruler138/CodeAnything-1.835M
Code: yuweiyang-anu/Code_Anything, branch release/sft-eval-history-20260910

Access is gated; request access on the repository page and authenticate with a Hugging Face user token after approval.

Downloads last month: -

Model tree for Ruler138/CodeAnything-Qwen3.5-9B-SFT-1.835M

Base model

Qwen/Qwen3.5-9B-Base

Finetuned

Qwen/Qwen3.5-9B

Adapter

(1004)

this model