Rewrite model card: BigBang-style structure, expanded benchmark comparison, remove banner

README.md

@@ -8,50 +8,33 @@ tags:
 pipeline_tag: text-generation
 ---
 
-<div align="center">
-<img src="assets/fx-bio-banner.png" alt="Fx-Bio Banner" width="100%"/>
-
-<p>
-<a href="https://huggingface.co/endless-frontier/Fx-Bio"><img alt="Model" src="https://img.shields.io/badge/%F0%9F%A4%97%20Model-endless--frontier%2FFx--Bio-536af5"/></a>
-<a href="LICENSE"><img alt="License" src="https://img.shields.io/badge/License-Apache--2.0-blue.svg"/></a>
-<img alt="Base Model" src="https://img.shields.io/badge/Base-DeepSeek--V4--Flash-6b5ce7"/>
-<img alt="Context" src="https://img.shields.io/badge/Context-1M%20tokens-0d9488"/>
-</p>
-
-<p>English ⬇️ &nbsp;|&nbsp; <a href="#chinese">中文 ⬇️</a></p>
-</div>
-
-**Fx-Bio-0913** is a biomedical reasoning large language model post-trained on **DeepSeek-V4-Flash**, developed by The Endless Frontier lab. Model weights are openly released under **Apache-2.0**.
+# Fx-Bio
 
-## Highlights
+## Introduction
 
-- 🧬 **Domain-specialized**: post-trained on biomedical reasoning tasks (BioMni / BioMystery-related data) on top of DeepSeek-V4-Flash.
-- 📈 **Consistent gains over the base model** on biomedical reasoning benchmarks, especially on harder problems — e.g. **+5.8** on BioMysteryBench Human-Difficult (avg@5) and **+6.6** on BiominiBench pass@3 over DeepSeek-V4-Flash.
-- 📏 **1M-token context length**, inherited from the base model, for long biomedical documents and multi-hop evidence.
-- 🔓 **Fully open weights** (bf16 safetensors) under Apache-2.0, with Megatron-LM training arguments (`args.json`) included for reproducibility.
+Fx-Bio-0913 is a biomedical reasoning large language model post-trained on **DeepSeek-V4-Flash**, developed by The Endless Frontier lab. It is specialized for biological and biomedical research tasks — including gene-function puzzles, experimental reasoning, and multi-step evidence integration — through post-training on curated biomedical reasoning data (Biomni / BioMystery-related tasks). The model retains the 1M-token context length of its base model, and its weights are openly released under Apache-2.0.
 
-## Performance
+## Main Results
 
-Evaluated on **BioMysteryBench** and **BiominiBench** against the DeepSeek-V4-Flash base model and Opus-5 (Accuracy %, higher is better):
+Fx-Bio-0913 on two biomedical reasoning benchmarks, **BioMysteryBench** and **BiominiBench**, compared with the DeepSeek-V4-Flash base model and Opus-5. Post-training yields consistent gains over the base model across all metrics, with the largest improvements on harder problems (BioMysteryBench Human-Difficult, +5.8) and on pass@3 of BiominiBench (+6.6).
 
-<div align="center">
-<img src="assets/fx-bio-eval.png" alt="Fx-Bio Evaluation Results" width="100%"/>
-</div>
-
-**BioMysteryBench**
+<p align="center">
+  <img src="./assets/fx-bio-eval.png" alt="Fx-Bio-0913 evaluation results on BioMysteryBench and BiominiBench" width="100%">
+</p>
 
-| Metric | DeepSeek-V4-Flash | Opus-5 | **Fx-Bio-0913** |
-|---|---|---|---|
-| Human-Solvable avg@5 | 85.2 | 89.3 | **85.5** |
-| Human-Difficult avg@5 | 31.8 | 45.9 | **37.6** |
-| pass@5 | 82.2 | 86.7 | **86.7** |
+## Benchmark Results
 
-**BiominiBench**
+Comparison of Fx-Bio-0913 with representative frontier models on BioMysteryBench (avg@5 and pass@5) and BiominiBench (avg@3 and pass@3). All scores are Accuracy (%) from our internal evaluation pipeline; higher is better.
 
-| Metric | DeepSeek-V4-Flash | Opus-5 | **Fx-Bio-0913** |
-|---|---|---|---|
-| avg@3 | 71.0 | 79.1 | **77.9** |
-| pass@3 | 78.7 | 86.7 | **85.3** |
+| Benchmark | Opus<br>5 | GPT<br>5.6 | Gemini 3.8<br>Flash | GLM<br>5.3 | Qwen3.8 Flash<br>Next (1M) | DeepSeek V4.1<br>Flash | DeepSeek V4<br>Flash | Fx-Bio<br>0913 |
+|:--|--:|--:|--:|--:|--:|--:|--:|--:|
+| **BioMysteryBench** |  |  |  |  |  |  |  |  |
+| Human-Solvable (avg@5) | 89.3 | 85.5 | 88.8 | 84.7 | 86.9 | 89.0 | 85.2 | **85.5** |
+| Human-Difficult (avg@5) | 45.9 | 34.1 | 42.4 | 47.1 | 49.4 | 36.5 | 31.8 | **37.6** |
+| pass@5 | 86.7 | 85.6 | 84.4 | 88.9 | 86.7 | 85.6 | 82.2 | **86.7** |
+| **BiominiBench** |  |  |  |  |  |  |  |  |
+| avg@3 | 79.1 | 68.2 | 69.3 | 81.1 | 80.6 | 80.9 | 71.0 | **77.9** |
+| pass@3 | 86.7 | 78.6 | 78.5 | 87.2 | 88.9 | 85.9 | 78.7 | **85.3** |
 
 ## Model Overview
 
@@ -83,12 +66,12 @@ outputs = model.generate(**inputs, max_new_tokens=1024)
 print(tokenizer.decode(outputs[0], skip_special_tokens=True))
 ```
 
-> [!Note]
-> Requires a `transformers` version that supports the `DeepseekV4ForCausalLM` architecture. For production inference, a dedicated serving framework (e.g. vLLM / SGLang) is recommended.
+> [!Important]
+> Requires a `transformers` version that supports the `DeepseekV4ForCausalLM` architecture. For production workloads or high-throughput scenarios, a dedicated serving framework (e.g. SGLang or vLLM) with DeepSeek-V4 support is recommended.
 
 ## Training
 
-Fx-Bio is post-trained on biomedical data (BioMni / BioMystery-related tasks). The detailed training recipe and data composition will be described in an upcoming technical report.
+Fx-Bio is post-trained on biomedical reasoning data (Biomni / BioMystery-related tasks) with Megatron-LM. The detailed training recipe and data composition will be described in an upcoming technical report.
 
 `args.json` contains the original Megatron-LM launch arguments for reproducibility.
 
@@ -99,86 +82,3 @@ For research use only. Model outputs may contain errors or inaccuracies and must
 ## License
 
 [Apache-2.0](LICENSE)
-
----
-
-<h1 id="chinese">Fx-Bio (0913) · 中文介绍</h1>
-
-**Fx-Bio-0913** 是基于 **DeepSeek-V4-Flash** 进行后训练（post-training）得到的生物领域推理大语言模型，由 The Endless Frontier 实验室发布。模型权重以 **Apache-2.0** 协议开放。
-
-## 亮点
-
-- 🧬 **领域专精**：在 DeepSeek-V4-Flash 基座之上，使用生物医学推理任务数据（BioMni / BioMystery 相关）进行后训练。
-- 📈 **稳定超越基座**：在两个生物医学推理基准上全面超过 DeepSeek-V4-Flash，困难样本上提升尤为明显——BioMysteryBench Human-Difficult（avg@5）**+5.8**，BiominiBench pass@3 **+6.6**。
-- 📏 **1M 超长上下文**：继承基座模型的 1M token 上下文能力，适配长篇生物医学文献与多跳证据推理。
-- 🔓 **完全开放权重**：bf16 safetensors，Apache-2.0 协议，并附 Megatron-LM 训练启动参数（`args.json`）便于复现。
-
-## 评测结果
-
-在 **BioMysteryBench** 和 **BiominiBench** 两个生物医学推理基准上与 DeepSeek-V4-Flash 基座模型、Opus-5 的对比如下（Accuracy %，越高越好）：
-
-<div align="center">
-<img src="assets/fx-bio-eval.png" alt="Fx-Bio 评测结果" width="100%"/>
-</div>
-
-**BioMysteryBench**
-
-| 指标 | DeepSeek-V4-Flash | Opus-5 | **Fx-Bio-0913** |
-|---|---|---|---|
-| Human-Solvable avg@5 | 85.2 | 89.3 | **85.5** |
-| Human-Difficult avg@5 | 31.8 | 45.9 | **37.6** |
-| pass@5 | 82.2 | 86.7 | **86.7** |
-
-**BiominiBench**
-
-| 指标 | DeepSeek-V4-Flash | Opus-5 | **Fx-Bio-0913** |
-|---|---|---|---|
-| avg@3 | 71.0 | 79.1 | **77.9** |
-| pass@3 | 78.7 | 86.7 | **85.3** |
-
-## 模型信息
-
-| 项目 | 内容 |
-|---|---|
-| 基座模型 | DeepSeek-V4-Flash |
-| 模型类型 | MoE 因果语言模型（`DeepseekV4ForCausalLM`） |
-| 权重格式 | bf16 safetensors（114 个分片） |
-| 上下文长度 | 1M tokens |
-| checkpoint | checkpoint-265 |
-| 协议 | Apache-2.0 |
-
-## 使用方法
-
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-model_id = "endless-frontier/Fx-Bio"
-
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(
-    model_id,
-    torch_dtype="bfloat16",
-    device_map="auto",
-)
-
-inputs = tokenizer("<你的 prompt>", return_tensors="pt").to(model.device)
-outputs = model.generate(**inputs, max_new_tokens=1024)
-print(tokenizer.decode(outputs[0], skip_special_tokens=True))
-```
-
-> [!Note]
-> 需要支持 `DeepseekV4ForCausalLM` 架构的 `transformers` 版本。生产环境推理建议使用 vLLM / SGLang 等专用推理框架。
-
-## 训练
-
-模型在后训练阶段使用了生物医学领域数据（BioMni / BioMystery 相关任务数据）。详细训练配方与数据组成将在后续技术报告中公布。
-
-`args.json` 为训练时的 Megatron-LM 启动参数，供复现参考。
-
-## 免责声明
-
-本模型仅供科研用途。模型输出可能包含错误或不准确信息，不应直接用于临床诊断或医疗决策。
-
-## 许可证
-
-[Apache-2.0](LICENSE)