Merge remote-tracking branch 'origin/main' into alekseymalakhov11-upd…

…ate-ci-cd
turbo-llm · Aug 24, 2024 · 4d97861 · 4d97861
2 parents 57e42a7 + 1d999de
commit 4d97861
Show file tree

Hide file tree

Showing 70 changed files with 774 additions and 4,723 deletions.
diff --git a/Makefile b/Makefile
@@ -39,6 +39,9 @@ ifneq ($(CODE),)
 	unify --in-place --recursive $(CODE) tutorials
 endif
 
+bash-dev:
+	docker run -v $(shell pwd):/app --rm -it ${DEV_IMAGE} bash
+
 lock:
 	rm poetry.lock
 	poetry lock

diff --git a/README.md b/README.md
@@ -56,7 +56,6 @@ Turbo-Alignment supports a wide range of methods for model training and alignmen
 - **📏** Length
 - **🌀** Perplexity
 - **🌟** METEOR
-- **📐** RAGAS
 - **🔍** Retrieval Utility
 
 <a name="-how-to-use"></a>
@@ -86,7 +85,7 @@ Examples of datasets are available [here](docs/dataset_example.md).
   - [RSO](#-RSO-sampling)
 - [Common](#-common)
   - [Preprocess](#-preprocess-common)
-  - [Convert to base](#-convert-to-base-common)
+  - [Merge adapters to base](#-merge-adapters-to-base-common)
 
 <a name="-train"></a>
 # Train
@@ -170,7 +169,9 @@ To launch RAG:
 ## 🚀 Installation
 
 ### 📦 Python Package
-⌛️ in progress..
+```bash
+pip install turbo-alignment
+```
 
 ### 🛠️ From Source
 For the latest features before an official release:

diff --git a/docs/dataset_example.md b/docs/dataset_example.md
@@ -11,7 +11,7 @@
 - [Pair Preferences Dataset](#-pair-preferences-dataset)
 - [KTO Dataset](#-kto-dataset)
 - [Sampling Dataset](#-sampling-dataset)
-- [Multimodal Dataset ](#-multimodal-dataset) (⌛️ Work in progress...)
+- [Multimodal Dataset ](#-multimodal-dataset)
 - [Classification Dataset](#-classification-dataset)
 - [DPPO Dataset](#-ddpo-dataset) (⌛️ Work in progress...)
 
@@ -118,9 +118,45 @@ Example:
 
 <a name="-multimodal-dataset"></a>
 ## Multimodal Dataset
-⌛️ in progress..
 
+- `messages`: `list[MultimodalChatMessage]` — This is a sequence of messages that make up the chat history. Each `ChatMessage` includes:
+  - `role` - The participant's role in the conversation (e.g., `user` or `bot`).
+  - `type` – The type of modality (e.g., `text` or `image`)
+  - `content` - If the `type` is `text`, it's the textual content of the message. If it's `image`, it's the file path.
 
+Example:
+```json
+{
+  "id": "0",
+  "messages": [
+    {
+      "role": "system",
+      "type": "text",
+      "content": "You are a Multimodal AI assistant."
+    },
+    {
+      "role": "user",
+      "type": "image",
+      "content": "/path/to/cat.jpg"
+    },
+    {
+      "role": "user",
+      "type": "image",
+      "content": "/path/to/dog.jpg"
+    },
+    {
+      "role": "user",
+      "type": "text",
+      "content": "What's the difference between these two images?"
+    },
+    {
+      "role": "bot",
+      "type": "text",
+      "content": "The two images in question both feature animals, albeit of different species. The first image depicts a dog, which is generally perceived as an animal that elicits positive emotional responses. The second image features a cat, which is also regarded as an animal that evokes a positive emotional response."
+    }
+  ]
+}
+```
 
 
 <a name="-classification-dataset"></a>

diff --git a/pyproject.toml b/pyproject.toml
@@ -3,7 +3,7 @@ name = "turbo-alignment"
 packages = [
     { include = "turbo_alignment" },
 ]
-version = "0.1.0"
+version = "0.0.2"
 description = "turbo-alignment repository"
 authors = ["T Mega Alignment Team <[email protected]>" ]
 
@@ -44,7 +44,6 @@ pydantic ="^2.7.0"
 timm ="^0.9.7"
 opencv-python = "^4.10.0.84"
 langchain-huggingface = "^0.0.3"
-ragas = "^0.1.10"
 
 [tool.poetry.group.deepspeed.dependencies]
 accelerate = "0.27"

diff --git a/tests/cli/test_dpo_train.py b/tests/cli/test_dpo_train.py
@@ -12,7 +12,10 @@
 
 @pytest.mark.parametrize(
     'config_path',
-    [FIXTURES_PATH / 'configs/train/dpo/base.json'],
+    [
+        FIXTURES_PATH / 'configs/train/dpo/base.json',
+        FIXTURES_PATH / 'configs/train/dpo/simpo.json',
+    ],
 )
 def test_dpo_train(config_path: Path):
     result = runner.invoke(

diff --git a/tests/cli/test_multimodal_inference.py b/tests/cli/test_multimodal_inference.py
@@ -1,26 +1,26 @@
-# from pathlib import Path
+from pathlib import Path
 
-# import pytest
-# from typer.testing import CliRunner
+import pytest
+from typer.testing import CliRunner
 
-# from tests.constants import FIXTURES_PATH
-# from turbo_alignment.cli import app
-# from turbo_alignment.settings.pipelines.inference.multimodal import (
-#     MultimodalInferenceExperimentSettings,
-# )
+from tests.constants import FIXTURES_PATH
+from turbo_alignment.cli import app
+from turbo_alignment.settings.pipelines.inference.multimodal import (
+    MultimodalInferenceExperimentSettings,
+)
 
-# runner = CliRunner()
+runner = CliRunner()
 
 
-# @pytest.mark.parametrize(
-#     'config_path',
-#     [
-#         FIXTURES_PATH / 'configs/inference/multimodal/llama_llava_clip_pickle.json',
-#     ],
-# )
-# def test_multimodal_inference_mlp_with_preprocessing(config_path: Path):
-#     result = runner.invoke(
-#         app, ['inference_multimodal', '--inference_settings_path', str(config_path)], catch_exceptions=False
-#     )
-#     assert result.exit_code == 0
-#     assert MultimodalInferenceExperimentSettings.parse_file(config_path).save_path.is_dir()
+@pytest.mark.parametrize(
+    'config_path',
+    [
+        FIXTURES_PATH / 'configs/inference/multimodal/llama_llava_clip_pickle.json',
+    ],
+)
+def test_multimodal_inference_mlp_with_preprocessing(config_path: Path):
+    result = runner.invoke(
+        app, ['inference_multimodal', '--inference_settings_path', str(config_path)], catch_exceptions=False
+    )
+    assert result.exit_code == 0
+    assert MultimodalInferenceExperimentSettings.parse_file(config_path).save_path.is_dir()
diff --git a/tests/cli/test_sft.py b/tests/cli/test_sft.py
@@ -14,24 +14,10 @@
     'config_path',
     [
         FIXTURES_PATH / 'configs/train/sft/base.json',
+        FIXTURES_PATH / 'configs/train/sft/sft_with_rm_metric.json',
     ],
 )
 def test_sft_train(config_path: Path):
     result = runner.invoke(app, ['train_sft', '--experiment_settings_path', str(config_path)], catch_exceptions=False)
     assert result.exit_code == 0
     assert SftTrainExperimentSettings.parse_file(config_path).log_path.is_dir()
-
-
-@pytest.mark.parametrize(
-    'config_path',
-    [
-        FIXTURES_PATH / 'configs/train/sft/resume_from_checkpoint.json',
-    ],
-)
-def test_sft_from_checkpoint(config_path: Path):
-    result = runner.invoke(
-        app,
-        ['train_sft', '--experiment_settings_path', str(config_path)],
-    )
-    assert result.exit_code == 0
-    assert SftTrainExperimentSettings.parse_file(config_path).log_path.is_dir()
diff --git a/tests/fixtures/configs/inference/rag/base.json b/tests/fixtures/configs/inference/rag/base.json
@@ -41,10 +41,10 @@
             "num_beams": 1,
             "max_new_tokens": 10,
             "repetition_penalty": 1.2,
+            "stop_strings": "</RS>",
             "do_sample": false
           },
           "custom_settings": {
-            "generation_eos_token": "</RS>",
             "skip_special_tokens": false,
             "remove_prompt": false
           }

diff --git a/tests/fixtures/configs/train/dpo/base.json b/tests/fixtures/configs/train/dpo/base.json
@@ -56,10 +56,10 @@
         "generator_transformers_settings": {
             "num_beams": 1,
             "do_sample": false,
+            "stop_strings": "</RS>",
             "max_new_tokens": 8
         },
         "custom_generation_settings": {
-            "generation_eos_token": "</RS>",
             "skip_special_tokens": false
           },
         "dataset_settings": {

diff --git a/tests/fixtures/configs/train/dpo/simpo.json b/tests/fixtures/configs/train/dpo/simpo.json
@@ -0,0 +1,134 @@
+{
+    "train_dataset_settings": {
+        "sources": [
+            {
+                "name": "rm_preferences_test",
+                "records_path": "tests/fixtures/datasets/rm/train_preferences.jsonl",
+                "sample_rate": 1
+            }
+        ],
+        "chat_settings":{
+            "prompt_template": {
+                "role_tag_mapping": {
+                    "bot": "<bot>",
+                    "user": "<user>",
+                    "system": "<system>"
+                },
+                "prefix_template": "<RS>{role}",
+                "suffix_template": "</RS>"
+            },
+            "max_tokens_count": 120
+        },
+        "add_labels": true,
+        "dataset_type": "pair_preferences"
+    },
+    "val_dataset_settings": {
+        "sources": [
+            {
+                "name": "rm_preferences_test",
+                "records_path": "tests/fixtures/datasets/rm/val_preferences.jsonl",
+                "sample_rate": 1
+            }
+        ],
+        "chat_settings":{
+            "prompt_template": {
+                "role_tag_mapping": {
+                    "bot": "<bot>",
+                    "user": "<user>",
+                    "system": "<system>"
+                },
+                "prefix_template": "<RS>{role}",
+                "suffix_template": "</RS>"
+            },
+            "max_tokens_count": 120
+        },
+        "add_labels": true,
+        "dataset_type": "pair_preferences"
+    },
+    "model_settings": {
+        "model_path": "tests/fixtures/models/llama2_tiny",
+        "model_type": "causal",
+        "transformers_settings": {},
+        "adapter_path": "tests/fixtures/models/llama2_tiny_fine_tuned_with_adapters/trainer",
+        "is_trainable": true
+    },
+    "cherry_pick_settings": {
+        "generator_transformers_settings": {
+            "num_beams": 1,
+            "do_sample": false,
+            "stop_strings": "</RS>",
+            "max_new_tokens": 8
+        },
+        "custom_generation_settings": {
+            "skip_special_tokens": false
+          },
+        "dataset_settings": {
+            "sources": [
+                {
+                    "name": "chat_test",
+                    "records_path": "tests/fixtures/datasets/chat/train_chat.jsonl",
+                    "num_samples": 2
+                }
+            ],
+            "prompt_template": {
+                "role_tag_mapping": {
+                    "bot": "<bot>",
+                    "user": "<user>",
+                    "system": "<system>"
+                },
+                "prefix_template": "<RS>{role}",
+                "suffix_template": "</RS>"
+            },
+            "dataset_type": "chat",
+            "max_tokens_count": 150,
+            "only_answer_loss": true
+        },
+        "metric_settings": [
+            {
+                "type": "length",
+                "parameters": {"need_average": [true]}
+            },
+            {
+                "type": "kl",
+                "parameters": {
+                    "need_average": [true],
+                    "ref_logits_type": "sft"
+                }
+            }
+        ]
+    },
+    "tokenizer_settings": {},
+    "trainer_settings": {
+        "evaluation_strategy": "steps",
+        "per_device_train_batch_size": 2,
+        "per_device_eval_batch_size": 2,
+        "gradient_accumulation_steps": 2,
+        "eval_steps": 4,
+        "save_steps": 4,
+        "logging_steps": 1,
+        "learning_rate": 0.0003,
+        "num_train_epochs": 2,
+        "lr_scheduler_type": "cosine",
+        "warmup_steps": 2,
+        "fp16": false,
+        "bf16": false,
+        "optim": "adamw_torch",
+        "save_total_limit": 1,
+        "average_log_prob": true,
+        "loss_settings": {
+            "loss_type": "simpo"
+        },
+        "sync_ref_settings": {
+            "sync_ref_model": false
+        },
+        "use_ref_model": false,
+        "use_sft_model": true,
+        "no_cuda": true
+    },
+    "wandb_settings": {
+        "project_name": "alignment",
+        "run_name": "dpo",
+        "entity": "turbo-alignment"
+      },
+    "log_path": "test_dpo_llama_train_output"
+}
diff --git a/tests/fixtures/configs/train/kto/base.json b/tests/fixtures/configs/train/kto/base.json
@@ -54,10 +54,10 @@
         "generator_transformers_settings": {
             "num_beams": 1,
             "do_sample": false,
+            "stop_strings": "</RS>",
             "max_new_tokens": 8
         },
         "custom_generation_settings": {
-            "generation_eos_token": "</RS>",
             "skip_special_tokens": false
           },
         "dataset_settings": {