Qwen3.5-9B Numeric Atom Tokens 2-Epoch
This repository contains a research checkpoint trained by our group. It is not an official Qwen release. It is published for reproducibility and comparison in vector graphics generation experiments.
Summary
The two-epoch numeric atom-token model used to initialize Image2SVG GRPO.
- Task family: Text2SVG/Image2SVG
- Project role: Image2SVG GRPO initialization
- Training date: 20260809
Background and experiment
The project studies editable SVG generation rather than raster-only image generation. The July work established SFT controls; later work tested numeric coordinate tokens, hard-example training, tool-oriented SVG editing, data- flywheel student training, and render-based reinforcement learning.
The Image2SVG DINO-GRPO initialization; tokenization reduced SVG token volume by about 40.26%.
Technical description
Qwen3.5-9B with integer tokens 10-512, padded expanded vocabulary, white RGB 512, BF16 and packed two-epoch SFT. The tokenizer must be kept with the weights.
Training data
The training data are not redistributed in this repository. Depending on the branch, the data came from the early SVG SFT corpus, the hardpair mixture, the numeric-token SVG corpus, aligned Image2SVG records, VSF tool records, or the simplified Cosmos/FLUX data flywheel. The source manifests and image assets remain in the research workspace and may have separate licenses.
Intended use and limitations
Use this model for research on SVG generation, vectorization, and controlled graphics generation. It may produce invalid SVG, omit details, or generate semantically incorrect geometry. It has not been safety-tuned for production use. Text2SVG checkpoints expect the text/chat format used by their training branch. Image2SVG checkpoints expect an image-conditioned prompt. Decoding parameters and tokenizer compatibility materially affect results.
Loading
This is a Transformers-format checkpoint. Load it with the matching Qwen3.5 code and the tokenizer shipped here. Numeric atom-token models must not use the base Qwen tokenizer: the expanded integer vocabulary and padded embeddings are part of the model.
License and attribution
The research artifact is released under the terms selected for this private research archive, subject to the upstream Qwen3.5 license and the licenses of all source data. Users are responsible for checking upstream model, dataset, and image terms before redistribution or deployment. This private archive is not a public license grant.
Provenance
- Source path on the training server:
/mnt/shared-storage-user/mineru4s/wangchuang/lgt/outputs/qwen35_9b_numeric_vocab_white512_2e_2n8g_20260809_v1/numeric_10_512_posinit_white512_packed_2e/checkpoint-106068 - Release account:
Yukinonooo - Release date:
2026-09-30 - This card was generated from the versioned release manifest in the project workspace.