library_name: litert-lm
license: gemma
base_model: OpenYourMind/gemma-4-12B-it-abliterated-uncensored
language:
- en
- ja
tags: - gemma4
- litert-lm
- webgpu
- abliterated
- experimental
OpenYourMind Gemma 4 12B abliterated → LiteRT-LM (experimental)
This browser-oriented LiteRT-LM artifact is converted from the BF16 full checkpoint, not a GGUF, of OpenYourMind/gemma-4-12B-it-abliterated-uncensored at revision 36ff0ae39a0e9037f3d50399d73d593caadfb9a0. Its model.safetensors SHA-256 is 6503b6fa17c2d7187e03a29cb16b0a659af41e8c53b0e30f4e4c1c2ae12cfe50. Source tokenizer_config.json is 2,102 bytes, SHA-256 09e6222fd7049dae603d0f61a8b10af5a813f22dca10e6cdc6918f76d7661104.
LiteRT Torch 0.10 used dynamic_wi4b32_afp32, a 4,096-token cache and prefill length 128. A separate copy of the source safetensors header had legacy vision tensor names adapted for the exporter; BF16 tensors and constants were converted to FP16 for WebGPU, and the main model weights were externalized within the bundle. Quantization changes the representation and can change outputs. This artifact is not bit-identical to the source checkpoint.
File
hina-12b-abliterated-block32-web-f16-external.litertlm
Size: 7,273,467,408 bytes
SHA-256: 5eae9001b422dbbaa85846926f97b4a91e634f17ea52db243631602fd9663867
Known quality failures and testing
This model has not passed Japanese quality validation. On Chrome 156, Apple M4 / 32 GB, WebGPU and @litert-lm/core 0.19.0, the model loaded and completed a short two-turn cue-recall test, but made spelling errors. In separate fresh sessions it repeatedly failed to answer the Japanese question “日本の首都はどこですか？”: greedy output discussed East Asia and temperature 1 / top-p 0.95 / top-k 64 output incorrectly called Japan the capital. A Japanese system-free prompt also drifted to the Sea of Japan. Greedy 2+2 returned 4 and an English capital question returned Tokyo. These are a few prompts, not a comprehensive evaluation.
With thinking enabled, one 64-token test produced a thought channel but no final answer. Logs also included external-weight streaming warnings and a WebGPU wait retry. Conversation deletion and engine disposal succeeded. Long-term stability, broad quality, other browsers and devices, and multimodal inference remain untested. This build is an experimental choice; do not assume factual accuracy.
As a limited comparison, a community Q4_K_M GGUF identified by its publisher as derived from the same OpenYourMind model also answered the Japanese capital question with a description of Japan as an East Asian island country, while correctly answering 2+2 and the English capital question. It was tested with llama.cpp 0.8.0, the model's chat template, temperature 0, thinking off. The GGUF publisher does not state the original BF16 revision. The GGUF and this LiteRT artifact use different quantizations, so this comparison does not establish bitwise equivalence or the precise cause of the failure.
License and provenance
The upstream model card labels the license gemma; this card mirrors that field. The Google Gemma 4 base is described as Apache-2.0 in its model card and license. Consult the upstream model card and applicable terms when using or redistributing derivatives. See SOURCE.txt for fixed hashes and conversion notes.