CoolFace
Modelpublic

Leven999/Qwen3.5-0.8B-ChemDSL-Distilled

sourceHugging Faceupdated 13d agoView on Hugging Face
1likes91downloads
Model Card

Qwen3.5-0.8B ChemDSL Distilled

Public model and browser deployment artifacts, published at the repository owner's request.

Repository layout

  • —Root: original merged BF16 checkpoint, tokenizer and configuration (unchanged).
  • —`onnx-wasm/`: explicit-state ONNX export, browser WASM inference, XGrammar constrained decoding, frozen Python/RDKit preprocessing, Node runtime, browser dependencies, conversion scripts and original validation evidence.
  • —`onnx-wasm/README_PUBLIC.md`: public deployment instructions.
  • —`PUBLIC_RELEASE_MANIFEST.json`: per-file hashes and exclusions.
  • —PUBLIC_UPLOAD_COMPLETE.json: appears only after every planned file has been verified on the Hub. Its absence means the incremental upload has not yet been fully verified.

The student is Qwen3.5-0.8B, sequence-distilled from the project's Qwen3.8-27B ChemDSL CFHH SFT-v3.1 teacher. Input is a complete mapped reaction SMILES, including product endpoints. This is not reactants-only product prediction.

Choose the correct model

For browser deployment use `onnx-wasm/model/decoder_fp16packed.onnx` together with `decoder_fp16packed.data` and `decoder_fp16packed.json`. Use onnx-wasm/web/engine.mjs; this is not an arbitrary Transformers.js pipeline export.

decoder_fp32 and decoder_fp16storage are conversion/reference variants. `decoder_q8` is a rejected INT8 experiment (33/61 top-1 agreement on probes), retained only as negative evidence. Do not use it as the released model.

The root checkpoint preserves its original tensor names and configuration. Turnkey AutoModelForCausalLM.from_pretrained compatibility has not been validated. It uses model.language_model.* tensor keys, a qwen3_5_text configuration, and records Transformers 5.14.1 with use_cache=false.

Measured limits

Default packed ONNX matched the original reference top-1 on 61/61 tested positions (maximum logit error approximately 9.33e-5). A synthetic browser run completed with 441 output tokens at approximately 1.4 output tokens/s on the tested CPU/WASM setup. These are conversion/smoke tests, not dataset-level accuracy evaluations. Original H100 throughput and accuracy must not be attributed to this browser export.

Constrained decoding is external code, not part of model weights. The bundled Python/RDKit stage prepares the input-conditioned schema and tokenized prompt; model execution and token masking then run locally in browser WASM. The proprietary owner compiler is not included or run by this demo. JSON/schema success is not proof of chemistry, mechanism, or executable-program correctness.

Provenance and terms

Original weight SHA-256: 79ea420139d48813dcd3cac6875142aba2b0909a9cd1da3d762e7c8282ce6f67. Root `ARTIFACTS.sha256` covers the six original checkpoint files. This publication does not modify or retrain them.

Upstream component licenses/notices are retained. Public accessibility does not replace the respective model, project-source and third-party license terms; no new blanket project license is assigned by this upload. See `onnx-wasm/THIRD_PARTY_NOTICES.md`.