- Restore vision_config block; clean eos_token_id to [1, 106]
- Align generation_config.json transformers_version field
Preserves local tokenizer_config.json (model_specific_special_tokens embedded into model during LEK training).
fix(config): sync with upstream Google Gemma 4 day-1 fixes (b446025, b4a6011)
- Restore vision_config block; clean eos_token_id to [1, 106]
- Align generation_config.json transformers_version field
Preserves local tokenizer_config.json (model_specific_special_tokens embedded into model during LEK training).
docs: correct base_model lineage for HF model tree
Snider+1: Virgil
0e862eb
HF uses base_model + base_model_relation frontmatter to rank models in
search results and render the model tree widget. The Lemma family's
true lineage is:
google/gemma-4-*-it
└── LetheanNetwork/<m> (finetune — our namespace fork)
└── lthn/<m> (finetune — LEK merged into weights)
└── lthn/<m>-mlx (quantized — mlx 4/8bit/bf16)
Previously this repo had base_model_relation set to quantized, which
was wrong — LEK merging is a finetune, not a quant. Fixing so the
model tree widget ranks the family correctly.
Co-Authored-By: Virgil <virgil@lethean.io>
First clean commit after orphan rewrite — previous history was inherited
from a mis-cloned lthn/lemer main repo and referenced LFS blobs that
don't exist in this sibling repo's storage.
This repo contains:
- Gemma 4 E2B with Lethean Ethical Kernel (LEK) merged into 100 text
attention projections (q/k/v/o_proj) via LoRA finetune, then folded
into the base weights
- Native MLX Q4 conversion via mlx_vlm.convert
- Full multimodal support (text, image, audio) with vision and audio
towers preserved unmodified from Google's upstream Gemma 4 E2B
- Generation config with Google's calibrated sampling defaults
(temperature=1.0, top_p=0.95, top_k=64)
- Model card documenting variant-specific sizing, sibling repos, MLX
Quick Start (mlx-lm, mlx-vlm, mlx_vlm.server), EUPL-1.2 licensing
Base fork at LetheanNetwork/lemer remains the unmodified Google fork
for users who want raw bf16 without the LEK shift. Full Lemma family
documentation at https://huggingface.co/lthn/lemer.
Co-Authored-By:…
