Model Zoo

Weights are stored in FP16 precision by default. Only TinyLlama models have variants of FP32 and FP16.

Naming convention: {P}?{F}?{E}?-{base_model}-epoch{1|2|3|5}-{fp16}?
Legend: P - Pretrained; F - Finetuned; E - Edited

Llama3.2 8B

P-llama3.2-8b-epoch1-fp16
PF-llama3.2-8b-epoch2-fp16
- Base model for finetuning: P-llama3.2-8b-epoch1
PFE-llama3.2-8b-epoch2-fp16
- Base model for AlphaEdit: PF-llama3.2-8b-epoch2

Llama3.2 3B

P-llama3.2-3b-epoch1-fp16
F-llama3.2-3b-epoch2-fp16
PF-llama3.2-3b-epoch2-fp16
- Base model for finetuning: P-llama3.2-3b-epoch1
PFE-llama3.2-3b-epoch2-fp16
- Base model for AlphaEdit: PF-llama3.2-3b-epoch2

TinyLlama

P-tinyllama-epoch1-fp16
F-tinyllama-epoch3-fp16
PF-tinyllama-epoch3-fp32
- Base model for finetuning: P-tinyllama-epoch1
PFE-tinyllama-epoch3-fp32
- Base model for AlphaEdit: PF-tinyllama-epoch3
PFE-tinyllama-epoch3-fp16
- Same as PFE-tinyllama-epoch3, but instead of FP32 precision, halved to FP16
PFE-tinyllama-epoch5-fp32
- Base model for AlphaEdit: PF-tinyllama-epoch5 (not ported)

Quantised TinyLlama (.GGUF format)