Improve language tag

README.md

@@ -1,214 +1,226 @@
----
-base_model: Qwen/Qwen2.5-7B
-library_name: peft
-language:
-- en
-license: agpl-3.0
-datasets:
-- OramaSearch/nlp-to-query-small
----
-
-# Query Translator Mini
-
-This repository contains a fine-tuned version of Qwen 2.5 7B model specialized in translating natural language queries into structured Orama search queries.
-
-The model uses PEFT with LoRA to maintain efficiency while achieving high performance.
-
-## Model Details
-
-### Model Description
-
-The Query Translator Mini model is designed to convert natural language queries into structured JSON queries compatible with the Orama search engine.
-
-It understands various data types and query operators, making it versatile for different search scenarios.
-
-### Key Features
-
-- Translates natural language to structured Orama queries
-- Supports multiple field types: string, number, boolean, enum, and arrays
-- Handles complex query operators: `gt`, `gte`, `lt`, `lte`, `eq`, `between`, `containsAll`
-- Supports nested properties with dot notation
-- Works with both full-text search and filtered queries
-
-## Usage
-
-```python
-import json, torch
-from peft import PeftModel
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-SYSTEM_PROMPT = """
-You are a tool used to generate synthetic data of Orama queries. Orama is a full-text, vector, and hybrid search engine.
-
-Let me show you what you need to do with some examples.
-
-Example:
-  - Query: `"What are the red wines that cost less than 20 dollars?"`
-  - Schema: `{ "name": "string", "content": "string", "price": "number", "tags": "enum[]" }`
-  - Generated query: `{ "term": "", "where": { "tags": { "containsAll": ["red", "wine"] }, "price": { "lt": 20 } } }`
-
-Another example:
-  - Query: `"Show me 5 prosecco wines good for aperitif"`
-  - Schema: `{ "name": "string", "content": "string", "price": "number", "tags": "enum[]" }`
-  - Generated query: `{ "term": "prosecco aperitif", "limit": 5 }`
-
-One last example:
-  - Query: `"Show me some wine reviews with a score greater than 4.5 and less than 5.0."`
-  - Schema: `{ "title": "string", "content": "string", "reviews": { "score": "number", "text": "string" } }]`
-  - Generated query: `{ "term": "", "where": { "reviews.score": { "between": [4.5, 5.0] } } }`
-
-The rules to generate the query are:
-
-- Never use an "embedding" field in the schema.
-- Every query has a "term" field that is a string. It represents the full-text search terms. Can be empty (will match all documents).
-- You can use a "where" field that is an object. It represents the filters to apply to the documents. Its keys and values depend on the schema of the database:
-  - If the field is a "string", you should not use operators. Example: `{ "where": { "title": "champagne" } }`.
-  - If the field is a "number", you can use the following operators: "gt", "gte", "lt", "lte", "eq", "between". Example: `{ "where": { "price": { "between": [20, 100] } } }`. Another example: `{ "where": { "price": { "lt": 20 } } }`.
-  - If the field is an "enum", you can use the following operators: "eq", "in", "nin". Example: `{ "where": { "tags": { "containsAll": ["red", "wine"] } } }`.
-  - If the field is an "string[]", it's gonna be just like the "string" field, but you can use an array of values. Example: `{ "where": { "title": ["champagne", "montagne"] } }`.
-  - If the field is a "boolean", you can use the following operators: "eq". Example: `{ "where": { "isAvailable": true } }`. Another example: `{ "where": { "isAvailable": false } }`.
-  - If the field is a "enum[]", you can use the following operators: "containsAll". Example: `{ "where": { "tags": { "containsAll": ["red", "wine"] } } }`.
-  - Nested properties are supported. Just translate them into dot notation. Example: `{ "where": { "author.name": "John" } }`.
-  - Array of numbers are not supported.
-  - Array of booleans are not supported.
-
-Return just a JSON object, nothing more.
-"""
-
-QUERY = "Show me some wine reviews with a score greater than 4.5 and less than 5.0."
-
-SCHEMA = {
-    "title": "string",
-    "description": "string",
-    "price": "number",
-}
-
-base_model_name = "Qwen/Qwen2.5-7B"
-adapter_path = "OramaSearch/query-translator-mini"
-
-print("Loading tokenizer...")
-tokenizer = AutoTokenizer.from_pretrained(base_model_name)
-
-print("Loading base model...")
-model = AutoModelForCausalLM.from_pretrained(
-    base_model_name,
-    torch_dtype=torch.float16,
-    device_map="auto",
-    trust_remote_code=True,
-)
-
-print("Loading fine-tuned adapter...")
-model = PeftModel.from_pretrained(model, adapter_path)
-
-if torch.cuda.is_available():
-    model = model.cuda()
-    print(f"GPU memory after loading: {torch.cuda.memory_allocated(0) / 1024**2:.2f} MB")
-
-messages = [
-    {"role": "system", "content": SYSTEM_PROMPT},
-    {"role": "user", "content": f"Query: {QUERY}\nSchema: {json.dumps(SCHEMA)}"},
-]
-
-prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-outputs = model.generate(
-    **inputs,
-    max_new_tokens=512,
-    do_sample=True,
-    temperature=0.1,
-    top_p=0.9,
-    num_return_sequences=1,
-)
-
-response = tokenizer.decode(outputs[0], skip_special_tokens=True)
-print(response)
-```
-
-## Training Details
-
-The model was trained on a NVIDIA H100 SXM using the following configuration:
-
-- Base Model: Qwen 2.5 7B
-- Training Method: LoRA
-- Quantization: 4-bit quantization using bitsandbytes
-- LoRA Configuration:
-  - Rank: 16
-  - Alpha: 32
-  - Dropout: 0.1
-  - Target Modules: Attention layers and MLP
-
-- Training Arguments:
-  - Epochs: 3
-  - Batch Size: 2
-  - Learning Rate: 5e-5
-  - Gradient Accumulation Steps: 8
-  - FP16 Training: Enabled
-  - Gradient Checkpointing: Enabled
-
-## Supported Query Types
-
-The model can handle various types of queries including:
-
-1. Simple text search:
-
-```json
-{
-    "term": "prosecco aperitif",
-    "limit": 5
-}
-```
-
-2. Numeric range queries:
-
-```json
-{
-    "term": "",
-    "where": {
-        "price": {
-            "between": [20, 100]
-        }
-    }
-}
-```
-
-3. Tag-based filtering:
-
-```json
-{
-    "term": "",
-    "where": {
-        "tags": {
-            "containsAll": ["red", "wine"]
-        }
-    }
-}
-```
-
-## Limitations
-
-- Does not support array of numbers or booleans
-- Maximum input length is 1024 tokens
-- Embedding fields are not supported in the schema
-
-## Citation
-
-If you use this model in your research, please cite:
-
-```
-@misc{query-translator-mini,
-  author = {OramaSearch Inc.},
-  title = {Query Translator Mini: Natural Language to Orama Query Translation},
-  year = {2024},
-  publisher = {HuggingFace},
-  journal = {HuggingFace Repository},
-  howpublished = {\url{https://huggingface.co/OramaSearch/query-translator-mini}}
-}
-```
-
-## License
-
-AGPLv3
-
-## Acknowledgments
-
+---
+base_model: Qwen/Qwen2.5-7B
+library_name: peft
+language:
+- zho
+- eng
+- fra
+- spa
+- por
+- deu
+- ita
+- rus
+- jpn
+- kor
+- vie
+- tha
+- ara
+license: agpl-3.0
+datasets:
+- OramaSearch/nlp-to-query-small
+---
+
+# Query Translator Mini
+
+This repository contains a fine-tuned version of Qwen 2.5 7B model specialized in translating natural language queries into structured Orama search queries.
+
+The model uses PEFT with LoRA to maintain efficiency while achieving high performance.
+
+## Model Details
+
+### Model Description
+
+The Query Translator Mini model is designed to convert natural language queries into structured JSON queries compatible with the Orama search engine.
+
+It understands various data types and query operators, making it versatile for different search scenarios.
+
+### Key Features
+
+- Translates natural language to structured Orama queries
+- Supports multiple field types: string, number, boolean, enum, and arrays
+- Handles complex query operators: `gt`, `gte`, `lt`, `lte`, `eq`, `between`, `containsAll`
+- Supports nested properties with dot notation
+- Works with both full-text search and filtered queries
+
+## Usage
+
+```python
+import json, torch
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+SYSTEM_PROMPT = """
+You are a tool used to generate synthetic data of Orama queries. Orama is a full-text, vector, and hybrid search engine.
+
+Let me show you what you need to do with some examples.
+
+Example:
+  - Query: `"What are the red wines that cost less than 20 dollars?"`
+  - Schema: `{ "name": "string", "content": "string", "price": "number", "tags": "enum[]" }`
+  - Generated query: `{ "term": "", "where": { "tags": { "containsAll": ["red", "wine"] }, "price": { "lt": 20 } } }`
+
+Another example:
+  - Query: `"Show me 5 prosecco wines good for aperitif"`
+  - Schema: `{ "name": "string", "content": "string", "price": "number", "tags": "enum[]" }`
+  - Generated query: `{ "term": "prosecco aperitif", "limit": 5 }`
+
+One last example:
+  - Query: `"Show me some wine reviews with a score greater than 4.5 and less than 5.0."`
+  - Schema: `{ "title": "string", "content": "string", "reviews": { "score": "number", "text": "string" } }]`
+  - Generated query: `{ "term": "", "where": { "reviews.score": { "between": [4.5, 5.0] } } }`
+
+The rules to generate the query are:
+
+- Never use an "embedding" field in the schema.
+- Every query has a "term" field that is a string. It represents the full-text search terms. Can be empty (will match all documents).
+- You can use a "where" field that is an object. It represents the filters to apply to the documents. Its keys and values depend on the schema of the database:
+  - If the field is a "string", you should not use operators. Example: `{ "where": { "title": "champagne" } }`.
+  - If the field is a "number", you can use the following operators: "gt", "gte", "lt", "lte", "eq", "between". Example: `{ "where": { "price": { "between": [20, 100] } } }`. Another example: `{ "where": { "price": { "lt": 20 } } }`.
+  - If the field is an "enum", you can use the following operators: "eq", "in", "nin". Example: `{ "where": { "tags": { "containsAll": ["red", "wine"] } } }`.
+  - If the field is an "string[]", it's gonna be just like the "string" field, but you can use an array of values. Example: `{ "where": { "title": ["champagne", "montagne"] } }`.
+  - If the field is a "boolean", you can use the following operators: "eq". Example: `{ "where": { "isAvailable": true } }`. Another example: `{ "where": { "isAvailable": false } }`.
+  - If the field is a "enum[]", you can use the following operators: "containsAll". Example: `{ "where": { "tags": { "containsAll": ["red", "wine"] } } }`.
+  - Nested properties are supported. Just translate them into dot notation. Example: `{ "where": { "author.name": "John" } }`.
+  - Array of numbers are not supported.
+  - Array of booleans are not supported.
+
+Return just a JSON object, nothing more.
+"""
+
+QUERY = "Show me some wine reviews with a score greater than 4.5 and less than 5.0."
+
+SCHEMA = {
+    "title": "string",
+    "description": "string",
+    "price": "number",
+}
+
+base_model_name = "Qwen/Qwen2.5-7B"
+adapter_path = "OramaSearch/query-translator-mini"
+
+print("Loading tokenizer...")
+tokenizer = AutoTokenizer.from_pretrained(base_model_name)
+
+print("Loading base model...")
+model = AutoModelForCausalLM.from_pretrained(
+    base_model_name,
+    torch_dtype=torch.float16,
+    device_map="auto",
+    trust_remote_code=True,
+)
+
+print("Loading fine-tuned adapter...")
+model = PeftModel.from_pretrained(model, adapter_path)
+
+if torch.cuda.is_available():
+    model = model.cuda()
+    print(f"GPU memory after loading: {torch.cuda.memory_allocated(0) / 1024**2:.2f} MB")
+
+messages = [
+    {"role": "system", "content": SYSTEM_PROMPT},
+    {"role": "user", "content": f"Query: {QUERY}\nSchema: {json.dumps(SCHEMA)}"},
+]
+
+prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+outputs = model.generate(
+    **inputs,
+    max_new_tokens=512,
+    do_sample=True,
+    temperature=0.1,
+    top_p=0.9,
+    num_return_sequences=1,
+)
+
+response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+print(response)
+```
+
+## Training Details
+
+The model was trained on a NVIDIA H100 SXM using the following configuration:
+
+- Base Model: Qwen 2.5 7B
+- Training Method: LoRA
+- Quantization: 4-bit quantization using bitsandbytes
+- LoRA Configuration:
+  - Rank: 16
+  - Alpha: 32
+  - Dropout: 0.1
+  - Target Modules: Attention layers and MLP
+
+- Training Arguments:
+  - Epochs: 3
+  - Batch Size: 2
+  - Learning Rate: 5e-5
+  - Gradient Accumulation Steps: 8
+  - FP16 Training: Enabled
+  - Gradient Checkpointing: Enabled
+
+## Supported Query Types
+
+The model can handle various types of queries including:
+
+1. Simple text search:
+
+```json
+{
+    "term": "prosecco aperitif",
+    "limit": 5
+}
+```
+
+2. Numeric range queries:
+
+```json
+{
+    "term": "",
+    "where": {
+        "price": {
+            "between": [20, 100]
+        }
+    }
+}
+```
+
+3. Tag-based filtering:
+
+```json
+{
+    "term": "",
+    "where": {
+        "tags": {
+            "containsAll": ["red", "wine"]
+        }
+    }
+}
+```
+
+## Limitations
+
+- Does not support array of numbers or booleans
+- Maximum input length is 1024 tokens
+- Embedding fields are not supported in the schema
+
+## Citation
+
+If you use this model in your research, please cite:
+
+```
+@misc{query-translator-mini,
+  author = {OramaSearch Inc.},
+  title = {Query Translator Mini: Natural Language to Orama Query Translation},
+  year = {2024},
+  publisher = {HuggingFace},
+  journal = {HuggingFace Repository},
+  howpublished = {\url{https://huggingface.co/OramaSearch/query-translator-mini}}
+}
+```
+
+## License
+
+AGPLv3
+
+## Acknowledgments
+
 This model builds upon the Qwen 2.5 7B model and uses techniques from the PEFT library. Special thanks to the teams behind these projects.