Add pipeline tag, paper link, and improve model card documentation

Hi! I'm Niels from the Hugging Face community science team. This PR improves the documentation for this model.

Key changes:
- Added `pipeline_tag: text-generation` to the YAML metadata for better discoverability.
- Added a link to the research paper: [Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction](https://huggingface.co/papers/2605.09649).
- Added a link to the official [GitHub repository](https://github.com/ngocbh/trimkv).
- Refined the "Quick Start" section to ensure users understand how to load the model using the `trimkv` library.

README.md

@@ -1,9 +1,10 @@
 ---
-license: apache-2.0
-datasets:
-- open-r1/OpenR1-Math-220k
 base_model:
 - Qwen/Qwen3-4B
+datasets:
+- open-r1/OpenR1-Math-220k
+license: apache-2.0
+pipeline_tag: text-generation
 tags:
 - math
 - trimkv
@@ -12,13 +13,18 @@ tags:
 - Compression
 ---
 
-> TRIM-KV is an efficient and learnable key–value eviction strategy designed to improve the efficiency of large language models (LLMs) in long-horizon inference.
+# TrimKV-Qwen3-4B-Math
+
+> **TRIM-KV** is an efficient and learnable key–value eviction strategy designed to improve the efficiency of large language models (LLMs) in long-horizon inference.
+
+This model is a Qwen3-4B variant fine-tuned with TRIM-KV on the `OpenR1-Math-220k` dataset. It is based on the research paper [Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction](https://huggingface.co/papers/2605.09649).
 
 The core idea behind TRIM-KV is to learn the intrinsic importance of each key–value pair at creation time, which we call *token retention*, and then decay this importance exponentially over time to mimic the standard inference running with eviction. 
 
 The retention score is query-agnostic and captures the long-term utility of tokens. This is different from attention scores, which are query-dependent: they capture the short-term utility for predicting the next token and are recomputed at every step, making them local, myopic, and highly dependent on the transient decoding state.
 
-<a href="https://arxiv.org/pdf/2512.03324"><img src="https://img.shields.io/badge/arxiv-2512.03324-red?style=for-the-badge"></a>
+- **Paper:** [Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction](https://huggingface.co/papers/2605.09649)
+- **Code:** [Official GitHub Repository](https://github.com/ngocbh/trimkv)
 
 ### Why TRIM-KV?
 
@@ -41,40 +47,21 @@ And it's interpretable
     <img width="1000" alt="teaser" src="https://github.com/ngocbh/trimkv/blob/main/assets/eviction.png?raw=true"/>
 </div>
 
-<div align="center">
-    <img width="1000" alt="teaser" src="https://github.com/ngocbh/trimkv/blob/main/assets/vis.png?raw=true"/>
-</div>
-
 ---
 
 ## Getting Started
 
-### Requirements
-
-- Python 3.11 or higher (tested with 3.12)
-- PyTorch 2.7.0 or higher (tested with 2.8.0)
-- FlashAttention 2.7.2.post1 or higher (tested with 2.8.0)
-- Transformers 4.57.1
-
-```sh
-pip install -r requirements.txt
-```
-
-This is a minimal set of requirements for training purposes. Additional dependencies may be needed for running specific experiments. We provided a full example of the environment used in our experiments in [`examples/env.yaml`](examples/env.yaml).
-
 ### Installation
 
-From the root of the repo:
+To use this model, you need to install the `trimkv` library from the [official repository](https://github.com/ngocbh/trimkv):
 
 ```sh
 git clone https://github.com/ngocbh/trimkv.git
 cd trimkv
 pip install -e .
-````
-
----
+```
 
-## Quick Start
+### Quick Start
 
 ```python
 import torch
@@ -82,7 +69,7 @@ from trimkv.models.qwen3 import TrimKVQwen3ForCausalLM
 from trimkv.cache_utils import TrimKVCache
 from transformers import AutoTokenizer
 
-model_path = "<TrimKV model_path here>"
+model_path = "ngocbh/TrimKV-Qwen3-4B-Math"
 download_from = "huggingface"  # options: "wandb", "local", "huggingface"
 
 model = TrimKVQwen3ForCausalLM.from_pretrained(
@@ -101,7 +88,7 @@ model.config.memory_size = 512
 model.config.buffer_size = 128
 
 tokenizer = AutoTokenizer.from_pretrained(
-    model.config.base_model,
+    "Qwen/Qwen3-4B",
     use_fast=True,
     padding_side="left",
 )
@@ -110,18 +97,15 @@ tokenizer = AutoTokenizer.from_pretrained(
 # Note: TRIM-KV uses TrimKVCache under the hood. So please pass TrimKVCache to model.generate
 ```
 
-For a runnable end-to-end example, see [`examples/test_qwen3.py`](examples/test_qwen3.py).
-
-## Released Models
+For a runnable end-to-end example, see [`examples/test_qwen3.py`](https://github.com/ngocbh/trimkv/blob/main/examples/test_qwen3.py).
 
-| Base Model                    | TRIM-KV Checkpoints                           | Training Datasets        | Training Context Len | Training $M$ |
-|------------------------------|-----------------------------------------------|--------------------------|-------------------------|--------------|
-| Qwen3-1.7B                     | [TRIM-KV-Qwen3-1.7B-Math](https://huggingface.co/ngocbh/TrimKV-Qwen3-1.7B-Math)            | OpenR1-Math-220k          | 16K   | 256     |
-| Qwen3-4B                     | [TRIM-KV-Qwen3-4B-Math](https://huggingface.co/ngocbh/TrimKV-Qwen3-4B-Math)            | OpenR1-Math-220k          | 16K   | 256     |
-| Qwen3-8B                     | [TRIM-KV-Qwen3-8B-Math](https://huggingface.co/ngocbh/TrimKV-Qwen3-8B-Math)            |  OpenR1-Math-220k          | 16K   | 256     |
-| Qwen3-14B                    | [TRIM-KV-Qwen3-14B-Math](https://huggingface.co/ngocbh/TrimKV-Qwen3-14B-Math)           |  OpenR1-Math-220k         | 16K   | 256     |
-| Qwen3-4B-Instruct-2507 | [TrimKV-Qwen3-4B-Instruct-2507](https://huggingface.co/ngocbh/TrimKV-Qwen3-4B-Instruct-2507) | Synth-Long, BookSum, Buddhi      |  128K      | 1024     |
-| Phi-3-mini-128k-instruct | [TrimKV-Phi-3-mini-128k-instruct](https://huggingface.co/ngocbh/TrimKV-Phi-3-mini-128k-instruct) | LongAlpaca          |  128K  | 512 |
-| DeepSeek-R1-Distill-Llama-8B                    | [TrimKV-DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/ngocbh/TrimKV-DeepSeek-R1-Distill-Llama-8B)           |  OpenR1-Math-220k         | 32K   | 256     |
+## Citation
 
----
+```bibtex
+@article{bui2025make,
+  title={Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Eviction},
+  author={Bui, Ngoc and Nguyen, Hieu Trung and Cohan, Arman and Ying, Rex},
+  journal={arXiv preprint arXiv:2512.03324},
+  year={2025}
+}
+```