add output for train loss in assertian err

tests/conftest.py

@@ -496,6 +496,12 @@ def dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff(
     return datasets.load_from_disk(ds_path)["train"]
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_llama_7m_model():
+    # download the model
+    return snapshot_download_w_retry("axolotl-ai-internal/llama-7m", repo_type="model")
+
+
 # # pylint: disable=redefined-outer-name,unused-argument
 # def test_load_fixtures(
 #     download_smollm2_135m_model,

tests/e2e/integrations/test_kd.py

@@ -90,7 +90,7 @@ class TestKnowledgeDistillation:
         train(cfg=cfg, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "model.safetensors").exists()
         check_tensorboard(
-            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/loss", 1.0, "Train loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -121,5 +121,5 @@ class TestKnowledgeDistillation:
         train(cfg=cfg, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.safetensors").exists()
         check_tensorboard(
-            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/loss", 1.0, "Train loss (%s) is too high"
         )

tests/e2e/multigpu/solo/test_flex.py

@@ -89,5 +89,5 @@ class TestPackedFlex:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train loss (%s) is too high"
         )

tests/e2e/multigpu/test_gemma3.py

@@ -96,5 +96,5 @@ class TestMultiGPUGemma3:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 1.8, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 1.8, "Train loss (%s) is too high"
         )

tests/e2e/multigpu/test_llama.py

@@ -43,7 +43,7 @@ class TestMultiGPULlama:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sequence_len": 2048,
                 "adapter": "lora",
                 "lora_r": 8,
@@ -94,7 +94,7 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -105,7 +105,7 @@ class TestMultiGPULlama:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sequence_len": 2048,
                 "sample_packing": True,
                 "eval_sample_packing": False,
@@ -159,14 +159,14 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     def test_dpo_lora_ddp(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sequence_len": 2048,
                 "sample_packing": False,
                 "eval_sample_packing": False,
@@ -244,7 +244,7 @@ class TestMultiGPULlama:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sequence_len": 2048,
                 "sample_packing": False,
                 "eval_sample_packing": False,
@@ -326,7 +326,7 @@ class TestMultiGPULlama:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "special_tokens": {
@@ -385,7 +385,7 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -396,7 +396,7 @@ class TestMultiGPULlama:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
                 "sequence_len": 1024,
@@ -457,7 +457,7 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     @require_torch_2_6_0
@@ -475,7 +475,7 @@ class TestMultiGPULlama:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
                 "sequence_len": 2048,
@@ -538,7 +538,7 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.1, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.1, "Train loss (%s) is too high"
         )
 
     def test_fsdp_qlora_prequant_packed(self, temp_dir):
@@ -618,7 +618,7 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -654,7 +654,7 @@ class TestMultiGPULlama:
             adapter = {}
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
                 "sequence_len": 1024,
@@ -702,7 +702,7 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -728,7 +728,7 @@ class TestMultiGPULlama:
             adapter = {}
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
                 "sequence_len": 1024,
@@ -776,7 +776,7 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -802,7 +802,7 @@ class TestMultiGPULlama:
             adapter = {}
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
                 "sequence_len": 1024,
@@ -850,7 +850,7 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     @pytest.mark.skip(
@@ -860,7 +860,7 @@ class TestMultiGPULlama:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "base_model": "axolotl-ai-internal/llama-7m",
                 "fix_untrained_tokens": True,
                 "sequence_len": 512,
                 "val_set_size": 0.0,
@@ -917,5 +917,5 @@ class TestMultiGPULlama:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 4.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 4.0, "Train loss (%s) is too high"
         )

tests/e2e/multigpu/test_ray.py

@@ -80,7 +80,7 @@ class TestMultiGPURay:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )
 
     @require_torch_lt_2_6_0
@@ -138,5 +138,5 @@ class TestMultiGPURay:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train loss (%s) is too high"
         )

tests/e2e/multigpu/test_sp.py

@@ -93,7 +93,7 @@ class TestSequenceParallelism:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.6, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.6, "Train loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(

tests/e2e/patched/test_fa_xentropy.py

@@ -86,5 +86,5 @@ class TestFAXentropyLlama:
         check_model_output_exists(temp_dir, cfg)
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 1.5, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 1.5, "Train loss (%s) is too high"
         )

tests/e2e/patched/test_unsloth_qlora.py

@@ -80,7 +80,7 @@ class TestUnslothQLoRA:
         check_model_output_exists(temp_dir, cfg)
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train loss (%s) is too high"
         )
 
     def test_unsloth_llama_qlora_unpacked(self, temp_dir):
@@ -130,7 +130,7 @@ class TestUnslothQLoRA:
         check_model_output_exists(temp_dir, cfg)
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -185,5 +185,5 @@ class TestUnslothQLoRA:
         check_model_output_exists(temp_dir, cfg)
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train loss (%s) is too high"
         )

tests/e2e/solo/test_flex.py

@@ -69,5 +69,5 @@ class TestPackedFlex(unittest.TestCase):
         train(cfg=cfg, dataset_meta=dataset_meta)
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train loss (%s) is too high"
         )

tests/e2e/test_llama_pretrain.py

@@ -84,5 +84,5 @@ class TestPretrainLlama:
             temp_dir + "/runs",
             "train/train_loss",
             loss_threshold,
-            "Train Loss is too high",
+            "Train Loss (%s) is too high",
         )

tests/e2e/test_packing_loss.py

@@ -68,5 +68,5 @@ class TestPackedLlama(unittest.TestCase):
         train(cfg=cfg, dataset_meta=dataset_meta)
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train loss (%s) is too high"
         )

tests/e2e/test_reward_model_smollm2.py

@@ -73,6 +73,6 @@ class TestRewardModelLoraSmolLM2(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.5, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.5, "Train loss (%s) is too high"
         )
         check_model_output_exists(temp_dir, cfg)

tests/test_perplexity.py

@@ -8,7 +8,7 @@ from transformers.models.auto.tokenization_auto import AutoTokenizer
 
 from axolotl.utils.callbacks.perplexity import Perplexity
 
-MODEL_NAME = "HuggingFaceTB/SmolLM2-135M"
+MODEL_NAME = "axolotl-ai-internal/llama-7m"
 
 
 @fixture()
@@ -36,7 +36,7 @@ One day, a little fish named Fin was swimming near the shore. He saw a big crab
 """
     result = metric.compute(model, [sample_text])
     ppl = result["score"]
-    assert round(ppl, 2) == 7.41
+    assert round(ppl, 2) == 75.14
 
 
 def test_perplexity_short(model, metric):
@@ -44,4 +44,4 @@ def test_perplexity_short(model, metric):
     sample_text = "Once upon a time, there was a little car named Beep. Beep loved to go fast and play in the sun."
     result = metric.compute(model, [sample_text])
     ppl = result["score"]
-    assert round(ppl, 2) == 10.33
+    assert round(ppl, 2) == 70.54