gemma 3

chat_templates/google/gemma-3-1b-it.jinja

@@ -27,9 +27,11 @@
 ' + (first_user_prefix if loop.first else "") }}
     {%- if message['content'] is string -%}
         {%- if message['role'] == 'assistant' -%}
-            {% generation %}{{ message['content'] | trim }}
+            {% generation %}{{ message['content'] | trim + '<end_of_turn>
+' }}{% endgeneration %}
         {%- else -%}
-            {{ message['content'] | trim }}
+            {{ message['content'] | trim + '<end_of_turn>
+' }}
         {%- endif -%}
     {# {%- elif message['content'] is iterable -%}
         {%- for item in message['content'] -%}
@@ -38,15 +40,10 @@
             {%- elif item['type'] == 'text' -%}
                 {{ item['text'] | trim }}
             {%- endif -%}
-        {%- endfor -%}
-    {%- else -%} #}
+        {%- endfor -%} #}
+    {%- else -%}
         {{ raise_exception("Invalid content type") }}
     {%- endif -%}
-    {{ '<end_of_turn>
-' }}
-    {%- if (message['role'] == 'assistant') -%}
-            {% endgeneration %}
-    {%- endif -%}
 {%- endfor -%}
 {%- if add_generation_prompt -%}
     {{'<start_of_turn>model

configs/context_numbers_10.yaml

@@ -0,0 +1,68 @@
+output_dir: "" # just a placeholder
+bf16: true
+model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+label_names: ["labels"]
+add_repeat_prompt: false
+add_negative_prompt: false
+# eval_on_start: True
+# eval_strategy: "steps"
+# eval_steps: 500
+# save_strategy: "no"
+# # save_steps: 500
+# logging_strategy: "steps"
+# logging_steps: 100
+# use_liger_kernel: true
+# remove_unused_columns: false
+
+# needed to avoid OOM by compute the metrics batch by batch
+# w/o this the trainer stores logits of all sample in memory...
+# batch_eval_metrics: true
+
+per_device_train_batch_size: 64
+per_device_eval_batch_size: 128
+max_new_tokens: 64
+gen_per_device_eval_batch_size: 128
+max_val_samples_per_ds: 500
+# optim: schedule_free_adamw
+learning_rate: 0.0001
+# lr_scheduler_type: "constant_with_warmup"
+neftune_noise_alpha: 1
+weight_decay: 0.01
+
+# LoRA
+lora_r: 16
+lora_dropout: 0.05
+target_modules:
+  - down_proj
+  - up_proj
+  - gate_proj
+
+# data
+train_ds_names:
+- data/raw_datasets/context_numbers_2
+- data/raw_datasets/context_numbers_3
+- data/raw_datasets/context_numbers_4
+- data/raw_datasets/context_numbers_5
+- data/raw_datasets/context_numbers_6
+- data/raw_datasets/context_numbers_7
+- data/raw_datasets/context_numbers_8
+- data/raw_datasets/context_numbers_9
+- data/raw_datasets/context_numbers_10
+
+val_ds_names:
+- data/raw_datasets/context_numbers_2
+- data/raw_datasets/context_numbers_3
+- data/raw_datasets/context_numbers_4
+- data/raw_datasets/context_numbers_5
+- data/raw_datasets/context_numbers_6
+- data/raw_datasets/context_numbers_7
+- data/raw_datasets/context_numbers_8
+- data/raw_datasets/context_numbers_9
+- data/raw_datasets/context_numbers_10
+
+test_ds_names:
+- data/raw_datasets/context_numbers_11
+- data/raw_datasets/context_numbers_12
+- data/raw_datasets/context_numbers_13
+- data/raw_datasets/context_numbers_14
+- data/raw_datasets/context_numbers_15