revert readme file

README.md

@@ -8,8 +8,7 @@ base_model:
 pipeline_tag: text-generation
 ---
 
-
-# Ring-lite-linear-preview
+# Ring-lite-distill-preview
 
 <p align="center">
     <img src="https://huggingface.co/inclusionAI/Ring-lite-distill-preview/resolve/main/ant-bailing.png" width="100"/>
@@ -21,20 +20,21 @@ pipeline_tag: text-generation
 
 ## Introduction
 
-Ring-lite-linear-preview is a hybrid-linear MoE LLM provided and open-sourced by InclusionAI, which has 17.1B parameters with 3.0B activated parameters. It is a long reasoning model based on hybrid-linear attention, achieving near-linear computational complexity and near-constant space complexity during inference. This model was converted from [Ling-lite-0220](https://huggingface.co/models/inclusionAI/Ling-lite), which adopts the softmax attention-based architecture. It matches the performance of DeepSeek-R1-Distill-Qwen-7B on standardized reasoning benchmarks while substantially reducing computational overhead in both training and inference phases. In certain generation speed tests based on vLLM, we observed that the throughput was more than doubled compared to softmax attention models of the same scale (e.g., Ling-lite). To the best of our knowledge, it is the first open-source hybrid-linear reasoning language model.
+Ring-lite-distill-preview is an MoE LLM provided and open-sourced by InclusionAI, which has 16.8B parameters with 2.75B activated parameters. It was fine-tuned from [Ling-lite](https://modelscope.cn/models/inclusionAI/Ling-lite) using extensive reasoning-focused instruction data. This model delivers performance comparable to DeepSeek-R1-Distill-Qwen-7B on reasoning benchmarks while achieving better results on general benchmarks, especially superior performance on function-calling evaluation benchmarks (e.g., TEval, BFCl_v2) and instruction-following benchmarks (e.g., IFEval). This demonstrates that Ring-lite-distill is a more balanced and versatile model. Additionaly, it maintains competitive latency and throughput compared to other reasoning LLMs of similar size.
+
 ## Model Downloads
 
 <div align="center">
 
 |     **Model**      | **#Total Params** | **#Activated Params** | **Context Length** | **Download** |
 | :----------------: | :---------------: | :-------------------: | :----------------: | :----------: |
-| Ring-lite-linear-preview |       17.1B       |         3.0B         |        64K         |      [🤗 HuggingFace](https://huggingface.co/inclusionAI/Ring-lite-distill)|
+| Ring-lite-distill-preview |       16.8B       |         2.75B         |        64K         |      [🤗 HuggingFace](https://huggingface.co/inclusionAI/Ring-lite-distill) |
 
 </div>
 
 ## Evaluation
-
-In terms of the evaluation of reasoning ability， Ring-lite-linear-preview achieves 55.0 on AIME24 and 93.8 on MATH-500.
+In order to fully evaluate the model's performance, we examined Ring-lite-distill-preview in terms of both reasoning ability and general ability.
+### Reasoning ability
 
 <div align="center">
 
@@ -42,38 +42,31 @@ In terms of the evaluation of reasoning ability， Ring-lite-linear-preview achi
 | :----------------: | :---------------: | :-------------------: | :----------------: | :----------: |
 | DeepSeek-R1-Distill-Qwen-7B (reported) |       55.5       |         92.8         |        49.1           |          37.6       |
 | DeepSeek-R1-Distill-Qwen-7B (reproduce)  |       53.2       |         93.7         |        50.4         |         36.5       |
-| Ring-lite-distill-preview-Stage-1 |      54.2     |         93.5       |         47.5       |       32.9      |
-| Ring-lite-linear-preview |      55.0     |         93.8       |         46.5       |       29.8      |
+| Ring-lite-distill-preview |       56.3       |         93.7         |        46.2        |        31.9       |
 
 </div>
 
-## Inference Speed
+### General ability
 
-To evaluate the generation throughput, we deploy Ring-lite-linear and the softmax-attention-based Ring-lite based on vLLM on a single NVIDIA A100 GPU. Specifically, the input sequence length is fixed to 1. The end-to-end (E2E) generation time required for generating output sequences of varying lengths is illustrated below. It is shown in the figure that at 32k output length, Ring-lite-linear-preview achieves 2.2× throughput of Ring-lite.
-
-<p align="center">
-    <img src="https://modelscope.cn/api/v1/models/inclusionAI/Ring-lite-linear-preview/repo?Revision=master&FilePath=throughput.png&View=true" width="600"/>
-<p>
-
-Additionally, to illustrate the advantage in inference speed, we present a comparison between Ring-lite-linear-preview and softmax-attention-based Ring-lite under a batch size of 64 and an output length of 16k (60x speedup). It can be observed that the KV cache usage of Ring-lite-linear-preview is nearly 1/6 that of Ring-lite, and the E2E time is reduced by 27.24% compared with Ring-lite.
-<p align="center">
-    <img src="https://modelscope.cn/api/v1/models/inclusionAI/Ring-lite-linear-preview/repo?Revision=master&FilePath=inference_speed.gif&View=true" width="600"/>
-<p>
+<div align="center">
 
-More details will be reported in our technical report [TBD]
+|     **Model**      | **IFEval**  | **T-eval** | **BFCL_v2** | **MMLU** |
+| :----------------: | :---------------: | :-------------------: | :----------------: | :----------: |
+| DeepSeek-R1-Distill-Qwen-7B (reproduce)  |       39.3       |         26.9          | 38.9 | 44.1 |
+| Ring-lite-distill-preview |      75.3 | 81.3 | 63.0 | 63.3 |
 
-## Requirements
-- [transformers](https://github.com/huggingface/transformers) >= 4.48.3
-- [flash-linear-attention](https://github.com/fla-org/flash-linear-attention) >= 0.2.1
+</div>
+More details will be reported in our technical report. [TBD]
 
 ## Quickstart
 
-Here is a code snippet to show you how to use the chat model with `modelscope`:
+### 🤗 Hugging Face Transformers
+Here is a code snippet to show you how to use the chat model with `transformers`:
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-model_name = "inclusionAI/Ring-lite-linear-preview"
+model_name = "inclusionAI/Ring-lite-distill-preview"
 
 model = AutoModelForCausalLM.from_pretrained(
     model_name,
@@ -105,17 +98,14 @@ generated_ids = [
 response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
 ```
 
-## Deployment
-
-Please refer to [Github](TBD)
-
 ## Dataset
+The training data of Ring-lite-distill-preview will be released soon. 
 
-The long reasoning sft data: [Ring-lite-distill-preview-sft-data](https://huggingface.co/datasets/inclusionAI/Ring-lite-distill-preview-sft-data)
-
+## Deployment
+Please refer to [GitHub](https://github.com/inclusionAI/Ring/blob/main/README.md)
 
 ## License
 This code repository is licensed under [the MIT License](https://huggingface.co/inclusionAI/Ring-lite-distill/blob/main/LICENSE).
 
 ## Citation
-[TBD]
+[TBD]