Llama-3-Kimodo-GGML

Llama-3-Kimodo-GGML

Native F32 GGML/GGUF LLM2Vec text encoder used by Kimodo, built from Meta Llama 3 and split into per-layer files.

What it is for

The repo holds the reusable LLM2Vec text encoder for Kimodo as F32 GGUF files: the embedding, one file per transformer layer, and the final norm. The layers stay split so kimodo.cpp can bound GPU memory use while it evaluates the encoder. It holds no diffusion model. Pair it with a matching Kimodo diffusion model, for example Kimodo-SMPLX-RP-v1-GGML. MANIFEST.json and SHA256SUMS record the source commits and the SHA-256 of each component.

Provenance

• Upstream sources: Meta Llama-3-8B-Instruct and the MIT-licensed McGill LLM2Vec MNTP and supervised adapters. Built with Meta Llama 3.

Use

From a kimodo.cpp checkout with the Hugging Face CLI installed, download the encoder and a diffusion model with:

scripts/download_gguf_weights.sh --output "$PWD"

Licence

The weights are a derivative of Meta Llama 3 and are under the Meta Llama 3 Community License, the licence of meta-llama/Meta-Llama-3-8B-Instruct. The McGill LLM2Vec adapters merged into them are MIT. LICENSE-META-LLAMA-3.txt holds the agreement and NOTICE the attribution it asks for; keep both with the files when you redistribute them, and follow the Acceptable Use Policy. Built with Meta Llama 3. The kimodo code repo is Apache 2.0; that licence covers the code, not these weights.