docs: expand release card format sections

README.md

@@ -1,38 +1,97 @@
 ---
-license: apache-2.0
-tags:
-  - text-classification
-  - intent-classification
-  - llama
-  - qwen
+
+base_model:
+    - Qwen/Qwen3-0.6B
+    - meta-llama/Llama-3.2-1B
 library_name: transformers
+pipeline_tag: text-classification
+tags:
+    - qwen3
+    - llama-3.2
+    - gguf
+    - onnx
+    - safetensors
+    - llama.cpp
+    - onnxruntime
+    - intent-classification
+    - text-classification
+license: apache-2.0
+
 ---
 
-# intent-classifier (release)
+# Intent Classifier (Release)
 
 This Hugging Face repo contains only the final released models for the
 intent-classifier project.
 
 It is intentionally limited to release artifacts:
-- merged full-weight model checkpoints,
-- GGUF exports for llama.cpp,
-- ONNX exports for runtime backends.
 
-Adapters and intermediate experiment artifacts are stored in:
-https://huggingface.co/kon172verma/intent-classifier-experiments
+- merged full-weight model checkpoints
+- GGUF exports for llama.cpp
+- ONNX exports for runtime backends.
 
-## Current model folders
+## Models
 
 - qwen3-0.6b
 - llama3.2-1b
 
-## Related repositories
+Both models are fine-tuned for intent classification and exported in multiple inference formats.
+
+- **Transformers / Safetensors:** Full-weight Hugging Face checkpoints for standard Transformers inference and downstream conversion.
+- **GGUF:** GGUF files are provided for llama.cpp inference.
+- **ONNX:** ONNX exports are provided for runtime backends.
+
+## Transformers / Safetensors
+
+The Transformers folders contain merged full-weight checkpoints in safetensors format.
+
+These are the canonical Hugging Face model artifacts for each selected release model and are the best starting point if you want to:
 
-- Training code: https://github.com/kon172verma/intent-classifier
-- Inference/benchmarking: https://github.com/kon172verma/intent-classifier-inference
-- Experiments (all adapters): https://huggingface.co/kon172verma/intent-classifier-experiments
+- run inference with Transformers,
+- inspect tokenizer and config files,
+- convert to another serving format,
+- fine-tune further from the released checkpoint.
+
+## GGUF
+
+The GGUF files are intended for inference with [llama.cpp](https://github.com/ggml-org/llama.cpp).
+
+Available quantization formats include:
+
+- Q4_K_M
+- Q6_K
+- Q8_0
+- F16
+
+## ONNX
+
+The ONNX folders contain exported model variants for ONNX Runtime backends.
+
+These artifacts are intended for deployment and benchmarking across runtimes such as CPU, CoreML, CUDA, or TensorRT pipelines, depending on the exported variant.
+
+When available, the ONNX exports may include multiple precision or quantization variants such as fp16 or int8.
+
+## Repository structure
+
+```text
+intent-classifier/
+├── qwen3-0.6b/
+│   ├── Transformers / Safetensors
+│   ├── GGUF
+│   └── ONNX
+│
+├── llama3.2-1b/
+│   ├── Transformers / Safetensors
+│   ├── GGUF
+│   └── ONNX
+│
+└── README.md
+```
+
+## Related repositories
 
-## Notes
+Training code and experiment artifacts are maintained separately.
 
-- Release selection and version mapping are tracked in the training repo.
-- This repo is kept minimal to make downstream inference and deployment workflows straightforward.
+- Training code: <https://github.com/kon172verma/intent-classifier>
+- Inference/benchmarking: <https://github.com/kon172verma/intent-classifier-inference>
+- Experiments (all adapters): <https://huggingface.co/kon172verma/intent-classifier-experiments>