fernandotonon/Llama-3-LLM2Vec-8B-Kimodo-TextEncoder-GGUF

Llama-3-LLM2Vec-8B-Kimodo-TextEncoder (GGUF, Q8_0)

Built with Meta Llama 3.

The text encoder of NVIDIA Kimodo packaged for QtMeshEditor, which runs Kimodo's text-to-motion natively (no Python) through its bundled llama.cpp.

What it is

Meta-Llama-3-8B-Instruct with the LLM2Vec MNTP and supervised LoRA adapters (McGill-NLP/LLM2Vec-Meta-Llama-3-8B-Instruct-mntp, -mntp-supervised, MIT) merged into its weights, saved as a headless bidirectional Llama (llama-embed architecture, attention.causal = false), converted with llama.cpp's convert_hf_to_gguf.py and quantized to Q8_0. Built by scripts/export-kimodo-text-encoder-gguf.py.

How Kimodo uses it

LLM2Vec embeddings with mean pooling over the prompt's own tokens: the input is <|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{prompt}<|eot_id|> and the embedding is the mean of the last hidden states of {prompt}<|eot_id|> (per-token embeddings, pooling = none, bidirectional attention). Measured against the original bf16 PyTorch encoder: cosine similarity 0.9999 on every test prompt.

Licence

This model is a derivative of Meta Llama 3 and is distributed under the Meta Llama 3 Community License — see LICENSE-Llama-3.txt. Use must comply with the Meta Llama 3 Acceptable Use Policy (USE_POLICY-Llama-3.md).

Meta Llama 3 is licensed under the Meta Llama 3 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved. (NOTICE)

The merged LLM2Vec adapters are MIT-licensed (LICENSE-LLM2Vec-MIT.txt, Copyright (c) 2024 McGill NLP).