File size: 2,802 Bytes
4fa885c 273c224 f1c28f5 f602de6 f1c28f5 f602de6 4fa885c 273c224 f602de6 273c224 f602de6 273c224 f602de6 273c224 f602de6 273c224 f602de6 273c224 f602de6 273c224 f602de6 273c224 f602de6 f1c28f5 f602de6 273c224 f602de6 273c224 f602de6 273c224 f602de6 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 | ---
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.
|