exo/lm-finetuner
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Batched loading and chunking of text datasets;

Browse Source
This commit is contained in:
extio-de
2025-02-09 16:17:27 +01:00
parent f253fde344
commit cb52446985
4 changed files with 24 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
testProject/config.ini
View File

@@ -25,9 +25,9 @@ locCustomPromptTemplate=./testProject/llama3_prompt_min.txt
# vInplace true: Validation will run after each batch of training epochs until vExpected has been reached, otherwise the next training batch starts
# vInplace false: Training is executed once for trEpochs epochs
# Optimizer schedule_free_*: Leave trEpochs blank
trEpochs=10
# optional int, maximum length of training sequence
trMaxSeqLength=
trEpochs=1
# int, maximum length of training sequence
trMaxSeqLength=2000
# optional scheduler type: linear, constant, cosine, polynomial, inverse_sqrt
trSchedulerType=
# optional optimizer; grokadamw, schedule_free_adamw, schedule_free_sgd
@@ -63,7 +63,7 @@ loraLayers=q_proj,k_proj
[Validation]
# enable validation
validate=true
validate=false
# Validate in place during training.
# In place validation: Validation will run after each batch of training epochs until vExpected has been reached. Requires more memory (both training and grader model are in memory at the same time)
# Normal validation: Validation is done once after the traning epochs have been finished