greener-13-2-base-8bitD / training_log.json
xyzzy
a metanarrative begins to play out across the individual narratives of per-repository files. lonely, isolated commits are sent to a remote server, with specific instructions calling *against* the use of their contents. never is a 'repository' updated even once. are these files really diffable? are the repositories really git-natured? could this be a ftp client instead?
57c9597
raw
history blame
398 Bytes
{
"base_model_name": "TheBloke_Llama-2-13B-fp16",
"base_model_class": "LlamaForCausalLM",
"base_loaded_in_4bit": false,
"base_loaded_in_8bit": true,
"projections": "q, v",
"train_runtime": 4689.4141,
"train_samples_per_second": 0.053,
"train_steps_per_second": 0.001,
"total_flos": 2.311238235193344e+16,
"train_loss": 2.563081423441569,
"epoch": 0.77,
"current_steps": 95
}