add finetome dataset to fixtures, check eval_loss in test (#2106) [skip ci]

* add finetome dataset to fixtures, check eval_loss in test

* add qwen 0.5b to pytest session fixture

tests/conftest.py

@@ -14,6 +14,12 @@ def download_smollm2_135m_model():
     snapshot_download("HuggingFaceTB/SmolLM2-135M")
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_qwen_2_5_half_billion_model():
+    # download the model
+    snapshot_download("Qwen/Qwen2.5-0.5B")
+
+
 @pytest.fixture(scope="session", autouse=True)
 def download_tatsu_lab_alpaca_dataset():
     # download the model
@@ -26,6 +32,11 @@ def download_mhenrichsen_alpaca_2k_dataset():
     snapshot_download("mhenrichsen/alpaca_2k_test", repo_type="dataset")
 
 
+def download_mlabonne_finetome_100k_dataset():
+    # download the model
+    snapshot_download("mlabonne/FineTome-100k", repo_type="dataset")
+
+
 @pytest.fixture
 def temp_dir():
     # Create a temporary directory

tests/e2e/multigpu/test_eval.py

@@ -7,10 +7,13 @@ from pathlib import Path
 
 import yaml
 from accelerate.test_utils import execute_subprocess_async
+from tbparse import SummaryReader
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
+from ..utils import most_recent_subdir
+
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
 
@@ -26,7 +29,7 @@ class TestMultiGPUEval:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "load_in_8bit": False,
                 "load_in_4bit": True,
                 "strict": False,
@@ -40,8 +43,8 @@ class TestMultiGPUEval:
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.1,
+                "val_set_size": 0.004,
-                "special_tokens": {"pad_token": "<|end_of_text|>"},
+                "special_tokens": {"pad_token": "<|endoftext|>"},
                 "datasets": [
                     {
                         "path": "teknium/GPT4-LLM-Cleaned",
@@ -66,6 +69,7 @@ class TestMultiGPUEval:
                 "saves_per_epoch": 1,
                 "logging_steps": 1,
                 "weight_decay": 0.0,
+                "use_tensorboard": True,
             }
         )
 
@@ -87,12 +91,18 @@ class TestMultiGPUEval:
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
+        tb_log_path = most_recent_subdir(temp_dir + "/runs")
+        event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
+        reader = SummaryReader(event_file)
+        df = reader.scalars  # pylint: disable=invalid-name
+        df = df[(df.tag == "eval/loss")]  # pylint: disable=invalid-name
+        assert df.value.values[-1] < 2.5, "Loss is too high"
 
     def test_eval(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "load_in_8bit": False,
                 "load_in_4bit": True,
                 "strict": False,
@@ -106,8 +116,8 @@ class TestMultiGPUEval:
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.1,
+                "val_set_size": 0.0004,
-                "special_tokens": {"pad_token": "<|end_of_text|>"},
+                "special_tokens": {"pad_token": "<|endoftext|>"},
                 "datasets": [
                     {
                         "path": "teknium/GPT4-LLM-Cleaned",
@@ -132,6 +142,7 @@ class TestMultiGPUEval:
                 "saves_per_epoch": 1,
                 "logging_steps": 1,
                 "weight_decay": 0.0,
+                "use_tensorboard": True,
             }
         )
 
@@ -153,3 +164,9 @@ class TestMultiGPUEval:
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
+        tb_log_path = most_recent_subdir(temp_dir + "/runs")
+        event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
+        reader = SummaryReader(event_file)
+        df = reader.scalars  # pylint: disable=invalid-name
+        df = df[(df.tag == "eval/loss")]  # pylint: disable=invalid-name
+        assert df.value.values[-1] < 2.9, "Loss is too high"