Instructions to use sharp-pixel/opensearch-agentic-dsl with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use sharp-pixel/opensearch-agentic-dsl with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("mistralai/Ministral-3-8B-Instruct-2512-BF16") model = PeftModel.from_pretrained(base_model, "sharp-pixel/opensearch-agentic-dsl") - Transformers
How to use sharp-pixel/opensearch-agentic-dsl with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="sharp-pixel/opensearch-agentic-dsl")# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("sharp-pixel/opensearch-agentic-dsl", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use sharp-pixel/opensearch-agentic-dsl with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "sharp-pixel/opensearch-agentic-dsl" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "sharp-pixel/opensearch-agentic-dsl", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/sharp-pixel/opensearch-agentic-dsl
- SGLang
How to use sharp-pixel/opensearch-agentic-dsl with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "sharp-pixel/opensearch-agentic-dsl" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "sharp-pixel/opensearch-agentic-dsl", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "sharp-pixel/opensearch-agentic-dsl" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "sharp-pixel/opensearch-agentic-dsl", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use sharp-pixel/opensearch-agentic-dsl with Docker Model Runner:
docker model run hf.co/sharp-pixel/opensearch-agentic-dsl
|
Download README.md from sharp-pixel/opensearch-agentic-dsl: direct link, hf CLI and curl.
- Browser
- Download file 3.72 kB
-
https://huggingface.co/sharp-pixel/opensearch-agentic-dsl/resolve/main/README.md
- Command line
-
hf download hf://sharp-pixel/opensearch-agentic-dsl/README.md
-
curl -L -o README.md https://huggingface.co/sharp-pixel/opensearch-agentic-dsl/resolve/main/README.md
3.72 kB
| base_model: mistralai/Ministral-3-8B-Instruct-2512-BF16 | |
| model_name: opensearch-agentic-dsl | |
| library_name: peft | |
| pipeline_tag: text-generation | |
| license: apache-2.0 | |
| language: | |
| - en | |
| tags: | |
| - base_model:adapter:mistralai/Ministral-3-8B-Instruct-2512-BF16 | |
| - lora | |
| - qlora | |
| - sft | |
| - transformers | |
| - trl | |
| - opensearch | |
| - text-to-dsl | |
| # opensearch-agentic-dsl | |
| A LoRA adapter that translates English shopper requests into OpenSearch DSL for | |
| a second-hand luxury catalog. It is designed for OpenSearch's native Agentic | |
| Search `QueryPlanningTool`. | |
| The adapter uses | |
| [mistralai/Ministral-3-8B-Instruct-2512-BF16](https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512-BF16) | |
| at revision `06cc81bfd6e45321d8fc8f816576c5b6ac67ec22`. Load both the base | |
| model and tokenizer at this revision, then attach the adapter with PEFT. | |
| ## Input and output | |
| Input combines shopper text, trusted application controls, the index mapping, | |
| and allowed query fields. The exact [system prompt](prompts/system.txt) and | |
| [user template](prompts/user.txt) are included. The | |
| [complete example](example.py) shows how to construct the three-line question | |
| and double JSON encode the mapping and field list. | |
| Output is one JSON request body containing `size`, integer `track_total_hits`, | |
| `query`, and optional `sort`. The planner applies required filters, price | |
| ceilings, and the selected ranking mode. Persona expansion contributes optional | |
| scoring. Index selection, search pipelines, and `_source` stay application-owned. | |
| ## Run the model | |
| Download `sharp-pixel/opensearch-agentic-dsl` and run the included example: | |
| ```bash | |
| hf download sharp-pixel/opensearch-agentic-dsl --local-dir opensearch-agentic-dsl | |
| cd opensearch-agentic-dsl | |
| python example.py | |
| ``` | |
| The example requires a CUDA GPU with enough memory for the BF16 base model and | |
| uses Transformers 5.13.0, PEFT 0.19.1, and PyTorch 2.13.0. It loads the bundled | |
| prompts and local adapter, uses greedy decoding, and prints the generated body. | |
| Set `ADAPTER_ID=sharp-pixel/opensearch-agentic-dsl` to load the adapter from the | |
| Hub instead. | |
| For vLLM, enable LoRA with maximum rank 16, load the language model only, and | |
| serve the adapter as `opensearch-agentic-dsl`. | |
| ## Training and evaluation | |
| Training used 4,640 synthetic examples and 1,160 held-out examples, grouped to | |
| keep related scenarios within a single split. Only the language-model attention | |
| and MLP projections were adapted; the vision encoder and projector stayed frozen. | |
| | Setting | Value | | |
| | --- | --- | | |
| | Method | QLoRA with completion-only supervision | | |
| | Quantization / compute | 4-bit NF4 / BF16 | | |
| | LoRA rank / alpha / dropout | 16 / 32 / 0.05 | | |
| | Epochs / sequence length | 2 / 2,112 tokens | | |
| | Learning rate / effective batch size | 0.0001 / 16 | | |
| | GPU / training time | RTX PRO 6000 Blackwell / 1.76 hours | | |
| | Metric | Value | | |
| | --- | ---: | | |
| | Training loss | 0.0038393672 | | |
| | Held-out loss | 0.0016606329 | | |
| | Held-out token accuracy | 99.9305% | | |
| Token accuracy is measured with teacher-forced targets, not freely generated | |
| requests. The synthetic evaluation does not establish real-world retrieval | |
| quality. Validate generated DSL against the actual mapping and application | |
| policy, and enforce OpenSearch permissions and resource limits. | |
| Training data, prompts, policy, and the dependency lockfile match this | |
| [verified source revision](https://github.com/sharp-pixel/LaTrouvaille/tree/7b26a2a56b71eab674ad8d53be5c1d06353967c8/query-understanding-training). | |
| ## License | |
| The adapter and accompanying model-card documentation are released under | |
| [Apache 2.0](LICENSE). The upstream model and tokenizer are also Apache 2.0; | |
| see [NOTICE](NOTICE) for attribution. Application source and training data retain | |
| their respective licenses. | |