---
license: apache-2.0
base_model: Qwen/Qwen3.5-4B
library_name: transformers
pipeline_tag: image-text-to-text
tags: [designcoder, tool-use, web-design, sft]
---

# DesignCoder 4B — AdamW step200

Full BF16 tool-using DesignCoder checkpoint, selected from
`xingxm/DesignCoder/designcoder_qwen3.5_4b_adamw_bs256_data41287_step200`.
This is an SFT initialization for DesignCoder RL. Its original weights, tokenizer,
processor and chat template are preserved byte-for-byte, without quantization or LoRA merging.

The source run uses Qwen3.5-4B, AdamW (learning rate 2e-5), global batch 256 and
41,287 DesignCoder v2 trajectories. The export is step200, epoch 1.5066.
The source model card records training on 8 nodes x 8 H20.

## Tool protocol and inference

Use the system prompt and tools in
[DesignCoder runtime](https://github.com/ZitengXue/DesignCoder/tree/main/examples/designcoder/runtime).
Landing: `design_search -> websearch -> HTML/CSS/JS`.
Dashboard: `design_search -> HTML/CSS/JS`.
Calls are written as `<tools>tool_name(...)</tools>`, followed by tool observations.
The final answer contains `<code lang="html">`, `<code lang="css">` and `<code lang="js">` blocks.

The original inference defaults are sampling with temperature 0.6, top_p 0.95,
top_k 20, repetition_penalty 1.0 and thinking enabled. Tool turns allow 1,536
new tokens; the final turn requests 32,768 tokens, capped by remaining room
in the 32,768-token context. Websearch requests four candidates per query.
Use `examples/designcoder/runtime/infer_designcoder.py` or the aligned RL agent;
a plain single-turn prompt does not reproduce the tool protocol.

## Evaluation provenance

The retained 200-case evaluation scores 83.2948 under the explicitly named
`fixed-assets-layout-v1` aggregation. The corresponding 4B Muon step200 scores
76.2119 under the same aggregation. The original judge identity is not recorded.
Older model-card numbers use other aggregation definitions and should not be mixed.
No new benchmark or GPU evaluation is claimed by this upload.

## Replacement and verification

This replaces the mistakenly selected Design.md-only AdamW step12588 checkpoint.
That previous model remains in Git history and the archive tag
`archive-designmd-4b-step12588-20261109`. The new root has one full
`model.safetensors`; it has no adapter or old weight shards.
`upload_manifest.json` records SHA256 values of every copied inference file.
`training_summary.json` records the checkpoint identity and source evidence.

```bash
hf download chjdl1816/designcoder-mc --local-dir models/designcoder-4b-adamw-step200
```

Use a clean local directory when switching from an earlier export.
The model architecture is `Qwen3_5ForConditionalGeneration`.
DesignEvaluator is a separate frozen 9B model and is not included here.