renaudb1999/le-harnais-ft-medium-regular
renaudb1999/le-harnais-ft-medium-regular is a 1 billion parameter ablation checkpoint based on Meta's Llama-3.2-1B-Instruct model. This model was trained on the Spider NL\u2192SQL dataset to evaluate the impact of JEPA (Joint Embedding Predictive Architecture) on loose view-alignment tasks. It specifically demonstrates that JEPA does not improve performance on NL\u2192SQL, serving as a negative result for research purposes rather than for inference.
Loading preview...
Model Overview
renaudb1999/le-harnais-ft-medium-regular is a 1 billion parameter ablation checkpoint derived from the meta-llama/Llama-3.2-1B-Instruct base model. This model is not intended for general inference but serves a specific research purpose within the le-harnais project.
Key Characteristics
- Base Model: Built upon
meta-llama/Llama-3.2-1B-Instruct, adhering to the Llama Community License. - Purpose: Designed as an ablation study to investigate the effectiveness of JEPA (Joint Embedding Predictive Architecture) on NL\u2192SQL tasks, specifically focusing on loose view-alignment.
- Training Data: Trained exclusively on the Spider NL\u2192SQL dataset.
- Research Finding: The primary outcome of this model is a negative result, indicating that JEPA does not provide a benefit for NL\u2192SQL tasks, with an execution match score of approximately 0.287.
Intended Use
This model is primarily for reproduction and continuation of the training study. It allows researchers to replicate the findings regarding JEPA's impact on NL\u2192SQL. For practical inference or real-world applications, users are directed to the main le-harnais-ft-agentworld-{1b,3b,8b} or le-harnais-ft-counsel models.
Available Formats
.safetensors: bf16 inference weights, compatible withtransformersorlh-serve/candle..Q4_K_M.gguf: Portable 4-bit quantized format, suitable forollama/llama.cppand Mac environments.