sadasat/librarian-c06

Librarian C06

An experimental, source-free Librarian derived from a custom Qwen3.5-0.8B checkpoint. It answers about Daniel's approved 27-record ideas library using merged model weights and a generic role prompt—not retrieval, a runtime adapter, source answers in the prompt, or server-side chat inference.

This is a narrow development result, not a general-purpose assistant or a held-out generalization result. Known questions guided the edits. The native development sweep reached 28/28 turns, 12/12 critical turns and 18/18 conversations. Five recorded browser rounds each reached those counts, but their 140 replies contain only 29 distinct case/text pairs. The additional 27/27 original-question probe includes wordings explicitly added to construction.

Files and integrity

• Root: the ordinary BF16 native checkpoint, original tokenizer/configuration assets and Apache 2.0 license. model.safetensors SHA-256: d63849518c46a19fd003ba2dc8b9e0f673422d57b3921da3ad9f5a430f22a18e.
• browser/: the canonical FP16/FP32 ONNX export, tokenizer, graph, provenance manifest, and integrity-bound 32 MiB download chunks of the same weights. Browser manifest SHA-256: 20e837655ff7cb03c5ed0ddbd4599ca685272797bedeab4f14c8c7c4d2c038bd.
• provenance/: reconstruction recipe, construction report and all-tensor audit. These are documentation, not inference inputs.
• release-files.json: byte lengths and SHA-256 checksums for every payload file.

Always use a full immutable Hub commit revision, not main, for a release. The native and ONNX representations have different storage dtypes and hashes; the browser manifest records their conversion lineage and tensor verification. The canonical ONNX external-data file and its chunks duplicate the same bytes; the browser downloads the chunks only, not both copies.

Run the browser application

The website repository supplies the browser application, pinned JavaScript/WASM runtime, static server and PiVo voice. Follow its operational guide. Both language-model inference and speech synthesis run on the visitor's CPU. This Hub repository supplies static downloads only; it is not an inference API. No Hugging Face token is required for public downloads.

Allow about 1.53 GB of model transfer plus runtime, several GB of free memory and tens of seconds per CPU reply. The measured real-model platform is Chrome for Testing 157 on Linux; physical-phone, Safari and universal device support are not established. The application keeps an explicit quality-limitations disclosure.

Native reproduction and tokenizer warning

Use the repository's pinned offline runtime and inference contract, not a default Hub pipeline. The original tokenizer asset is intentionally unchanged. The native reference uses explicit TokenizersBackend, fix_mistral_regex=False, Transformers 5.22.1 and tokenizers 0.23.3. Its effective backend SHA-256 is a57b5c62cd5f242a06c632b82cc0b4f8239d08cc80fd9347fc1f629be977c6ca. Silent tokenizer repair or a different chat template invalidates the recorded input-ID parity. Thinking is disabled, with 160 new tokens, temperature 0.7, top-p 0.9 and top-k 40. Browser and native execution are evaluated separately.

Method, modifications and limitations

The inherited custom base retains historical pruning/configuration and four embedding-coordinate edits. Relative to that base, C06 changes exactly one of 320 tensors: model.layers.23.mlp.down_proj.weight. Six gradient-free analytic readout projections fit 89 complete teaching units / 3,657 target positions. Role-aware protection excludes earlier assistant-answer targets that otherwise conflicted with the teaching objective, while retaining system/user and post-answer protection. This is example-supervised editing, not data-free learning.

A one-tensor change is not a tiny change: the relative writer-matrix update norm is 1.879. Sampled protected drift and a narrow collateral-control question do not prove broad capability preservation. Teacher-forced construction loss is not semantic accuracy. Novel wording, long conversations, unrelated knowledge and hallucination behavior require further evaluation.

The single study with graphs and complete evidence records successful and unsuccessful candidates, limitations, raw replies, external semantic judgments and literature, including arXiv:2611.09382. The recorded five-round browser result binds its original runtime bytes; publishing these identical model files does not itself constitute a new evaluation.

Derived from Qwen/Qwen3.5-0.8B revision 2fc06364715b967f1860aea9cf38778875588b17, under Apache 2.0. Modified checkpoint and browser conversion: Librarian research, November 2026. Retain the included license and this modification notice when redistributing.