Dataset is now directly tokenized to avoid further post-processing of the trainer; locCustomPromptTemplate to specify a custom Jinja template; showChatTemplate for debugging;

testProject/config.ini

@@ -3,6 +3,8 @@
 device=cuda:0
 # purge existing workdir, adapters and full output when the program starts
 purgeTargetDirectories=false
+# For debugging
+showChatTemplate=true
 
 [Trainer]
 # activate trainer
@@ -10,13 +12,15 @@ train=true
 # save lora adapter to disk after training. Also required for validation and merging
 storeAdapter=true
 # path to base model (hf safetensors repo)
-locBaseModel=../../lm/models/safetensors/Qwen2.5-1.5B-Instruct/
+locBaseModel=../../lm/models/safetensors/Llama-3.2-3B-Instruct-uncensored/
 # path to datasets
 locDataset=./testProject/dataset/
 # path: working directory
 locWorkdir=./testProject/work/
 # path: lora adapter output
 locAdapter=./testProject/adapter/
+# path: custom chat template (Jinja template), optional
+locCustomPromptTemplate=./testProject/llama3_prompt_min.txt
 # training epochs (how many times each dataset entry is trained)
 #   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

testProject/llama3_prompt_min.txt

@@ -0,0 +1,3 @@
+{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>
+
+'+ message['content'] | trim + '<|eot_id|>' %}{{ content }}{% endfor %}