Update README.md

README.md

@@ -14,9 +14,9 @@ tags:
   - pathogenicity
   - transmissibility
   - host-tropism
-  - hvue-v2
+  - hvue
 datasets:
-  - duttaprat/HVUE-v2
+  - duttaprat/HVUE
 pipeline_tag: feature-extraction
 widget:
   - text: "ATGCGTACGTTAGCCGATCG"
@@ -29,15 +29,12 @@ widget:
 
 [![Preprint](https://img.shields.io/badge/bioRxiv-2026-B31B1B)](https://www.biorxiv.org/content/10.64898/2026.03.18.712700v1)
 [![Code](https://img.shields.io/badge/Code-GitHub-black)](https://github.com/duttaprat/HViLM)
-[![Dataset](https://img.shields.io/badge/Dataset-HVUE--v2-yellow)](https://huggingface.co/datasets/duttaprat/HVUE-v2)
+[![Dataset](https://img.shields.io/badge/Dataset-HVUE-yellow)](https://huggingface.co/datasets/duttaprat/HVUE)
 [![License](https://img.shields.io/badge/License-Apache%202.0-green.svg)](LICENSE)
 
 </div>
 
-> [!IMPORTANT]
-> **HVUE v2 supersedes the original HVUE benchmark.**  
-> The original HVUE v1 benchmark contained substantial cross-split sequence similarity that could inflate held-out performance estimates. HVUE v2 was rebuilt using source-sequence clustering **before** train/validation/test assignment and chunking, followed by exact- and near-match leakage auditing.  
-> **Use [duttaprat/HVUE-v2](https://huggingface.co/datasets/duttaprat/HVUE-v2) for current HViLM evaluation.**
+
 
 ## Model Description
 
@@ -45,7 +42,7 @@ widget:
 
 The architecture and tokenizer remain those of DNABERT-2; continued pre-training updates the model weights using a masked-language-modeling objective on the virus-focused corpus.
 
-**Preprint:** *HViLM: A Foundation Model for Viral Genomics Enables Multi-Task Prediction of Pathogenicity, Transmissibility, and Host Tropism*  
+**Preprint:** *HViLM: A Foundation Model for Viral Genomics Enables Prediction of Pathogenicity, Transmissibility, and Host Tropism*  
 [bioRxiv 2026.03.18.712700](https://www.biorxiv.org/content/10.64898/2026.03.18.712700v1)
 
 **Authors:** Pratik Dutta, Jack Vaska, Pallavi Surana, Rekha Sathian, Max Chao, Zhihan Zhou, Han Liu, and Ramana V. Davuluri
@@ -56,7 +53,7 @@ The architecture and tokenizer remain those of DNABERT-2; continued pre-training
 
 ## HViLM Model Family
 
-HViLM-base is the continued-pretrained foundation model. Official task-specific models fine-tuned on HVUE v2 are released as standalone checkpoints:
+HViLM-base is the continued-pretrained foundation model. Official task-specific models fine-tuned on HVUE are released as standalone checkpoints:
 
 | Resource | Purpose |
 |---|---|
@@ -64,7 +61,7 @@ HViLM-base is the continued-pretrained foundation model. Official task-specific
 | [HViLM-Patho](https://huggingface.co/duttaprat/HViLM-Patho) | Pathogenicity classification |
 | [HViLM-R0](https://huggingface.co/duttaprat/HViLM-R0) | Transmissibility classification |
 | [HViLM-Tropism](https://huggingface.co/duttaprat/HViLM-Tropism) | Human host-tropism classification |
-| [HVUE-v2](https://huggingface.co/datasets/duttaprat/HVUE-v2) | Leakage-controlled benchmark |
+| [HVUE](https://huggingface.co/datasets/duttaprat/HVUE) | Leakage-controlled benchmark |
 
 The complete project is also grouped in the **HViLM: Human Virome Language Model** collection on the [duttaprat Collections page](https://huggingface.co/duttaprat/collections).
 
@@ -75,9 +72,9 @@ The complete project is also grouped in the **HViLM: Human Virome Language Model
 - **Virus-focused continued pre-training:** approximately 5M non-redundant fragments derived from VIRION-linked virus sequences.
 - **DNABERT-2 initialization:** preserves the DNABERT-2 architecture and BPE tokenizer while adapting model weights to virus sequence data.
 - **Three official downstream models:** pathogenicity, transmissibility, and host tropism.
-- **HVUE v2 evaluation:** cluster-aware splitting before chunking, with multiple similarity stringencies and sequence lengths.
+- **HVUE evaluation:** cluster-aware splitting before chunking, with multiple similarity stringencies and sequence lengths.
 - **Parameter-efficient downstream adaptation:** official task models were trained with LoRA.
-- **Public reproducibility resources:** base model, three task-specific checkpoints, HVUE v2 benchmark, and project code are released publicly.
+- **Public reproducibility resources:** base model, three task-specific checkpoints, HVUE benchmark, and project code are released publicly.
 
 ---
 
@@ -105,7 +102,7 @@ HViLM-base is derived from **DNABERT-2 (117M parameters)**.
 | Approximate training time | 72 hours |
 | Held-out MLM accuracy | 94.2% |
 
-**Sequence-length note:** HViLM uses BPE tokenization, so nucleotide length and model-token length are not equivalent. The continued-pretraining corpus used 1000-nt sequence fragments; downstream configurations are described by nucleotide length in HVUE v2.
+**Sequence-length note:** HViLM uses BPE tokenization, so nucleotide length and model-token length are not equivalent. The continued-pretraining corpus used 1000-nt sequence fragments; downstream configurations are described by nucleotide length in HVUE.
 
 ---
 
@@ -151,7 +148,7 @@ For sequence-level representations, pooling strategy should be chosen according
 
 ## Use the Official Fine-tuned Models
 
-If the goal is one of the three HVUE v2 tasks, users can load the corresponding task model directly; `HViLM-base` does not need to be loaded separately.
+If the goal is one of the three HVUE tasks, users can load the corresponding task model directly; `HViLM-base` does not need to be loaded separately.
 
 ### Pathogenicity
 
@@ -185,6 +182,7 @@ Labels:
 - `0` → `R0_LT_1`
 - `1` → `R0_GE_1`
 
+These label names are retained for checkpoint compatibility. The task is a binary classification of documented sustained human transmission capacity, informed by reported R₀ values rather than predicting R₀ itself.
 ### Host Tropism
 
 Use:
@@ -202,9 +200,9 @@ See the individual model cards for full usage examples and task-specific limitat
 
 ---
 
-## HVUE v2 Benchmark
+## HVUE Benchmark
 
-[HVUE v2](https://huggingface.co/datasets/duttaprat/HVUE-v2) is the current evaluation benchmark for HViLM. It replaces HVUE v1.
+[HVUE](https://huggingface.co/datasets/duttaprat/HVUE) is the current evaluation benchmark for HViLM. 
 
 The benchmark was reconstructed to reduce supervised train-test leakage using the following ordering:
 
@@ -213,9 +211,9 @@ The benchmark was reconstructed to reduce supervised train-test leakage using th
 3. assign complete clusters to train/validation/test splits;
 4. chunk sequences only **after** split assignment;
 5. remove exact duplicate chunks;
-6. audit cross-split exact and near matches.
+6. audit cross-split exact matches.
 
-HVUE v2 includes:
+HVUE includes:
 
 - **Pathogenicity**
 - **Transmissibility**
@@ -227,7 +225,7 @@ Across the benchmark, configurations evaluate different sequence lengths (500, 1
 
 The primary results below use the **standard 1000-nt configuration** for each task.
 
-| Task | HVUE v2 configuration | Accuracy | F1 | MCC | Official model |
+| Task | HVUE configuration | Accuracy | F1 | MCC | Official model |
 |---|---|---:|---:|---:|---|
 | Pathogenicity | `standard_capped_1000bp` | **92.39** | **91.32** | **83.10** | [HViLM-Patho](https://huggingface.co/duttaprat/HViLM-Patho) |
 | Transmissibility | `standard_capped_1000bp` | **87.50** | **86.16** | **72.66** | [HViLM-R0](https://huggingface.co/duttaprat/HViLM-R0) |
@@ -235,7 +233,7 @@ The primary results below use the **standard 1000-nt configuration** for each ta
 
 The directory/configuration identifiers retain `bp` for release stability; manuscript and descriptive text use **nt** for nucleotide sequence length.
 
-### Interpretation of the HVUE v2 Results
+### Interpretation of the HVUE Results
 
 Under leakage-controlled evaluation, the effect of virus-focused continued pre-training is **task dependent**:
 
@@ -243,9 +241,9 @@ Under leakage-controlled evaluation, the effect of virus-focused continued pre-t
 - **Transmissibility:** HViLM and DNABERT-2 are close (86.16 vs. 85.81 F1), and HViLM is essentially tied with DNABERT-MB (86.16 vs. 86.15 F1).
 - **Host Tropism:** HViLM shows the largest F1 improvement, reaching 74.49 compared with 64.82 for class-balanced DNABERT-2.
 
-These results support a more specific conclusion than the original HVUE v1 evaluation: virus-focused continued pre-training provides its clearest benefit on the more challenging Host Tropism task, while gains on Pathogenicity and Transmissibility are smaller.
+These results support a specific conclusion: virus-focused continued pre-training provides its clearest benefit on the more challenging Host Tropism task, while gains on Pathogenicity and Transmissibility are smaller.
 
-For complete baseline comparisons, hard-split evaluations, temporal evaluations, and sequence-length analyses, see the [HViLM GitHub repository](https://github.com/duttaprat/HViLM) and [HVUE-v2](https://huggingface.co/datasets/duttaprat/HVUE-v2).
+For complete baseline comparisons, hard-split evaluations, temporal evaluations, and sequence-length analyses, see the [HViLM GitHub repository](https://github.com/duttaprat/HViLM) and [HVUE](https://huggingface.co/datasets/duttaprat/HVUE).
 
 ---
 
@@ -277,11 +275,10 @@ These observations are **hypothesis-generating**. Sequence similarity between vi
 
 ## Limitations
 
-- HVUE v2 controls supervised split leakage through source-level clustering and auditing, but sequence-similarity thresholds cannot eliminate every form of biological relatedness.
+- HVUE controls supervised split leakage through source-level clustering and exact-match auditing, but sequence-similarity thresholds cannot eliminate every form of biological relatedness.
 - The complete historical training exposure of the original DNABERT-2 model cannot be reconstructed; therefore, absence of all possible ancestral pretraining exposure to benchmark-related sequences cannot be guaranteed.
 - Host association is biologically context-dependent and may include multi-host, zoonotic, and reverse-zoonotic relationships; the benchmark uses a simplified binary formulation.
-- R₀-based transmissibility labels simplify a continuous, context-dependent epidemiological quantity into a binary benchmark task.
-- Performance differences between closely matched models should not be interpreted as statistically meaningful without uncertainty estimates or repeated evaluations.
+- Transmissibility labels reduce a continuous, context-dependent epidemiological quantity to a binary benchmark task.- Performance differences between closely matched models should not be interpreted as statistically meaningful without uncertainty estimates or repeated evaluations.
 - Attention-based motif analyses should be considered exploratory rather than direct evidence of mechanism.
 - HViLM predictions are research outputs and are not intended to replace experimental, clinical, epidemiological, or public-health assessment.
 
@@ -293,7 +290,7 @@ These observations are **hypothesis-generating**. Sequence similarity between vi
 - **Pathogenicity model:** [duttaprat/HViLM-Patho](https://huggingface.co/duttaprat/HViLM-Patho)
 - **Transmissibility model:** [duttaprat/HViLM-R0](https://huggingface.co/duttaprat/HViLM-R0)
 - **Host Tropism model:** [duttaprat/HViLM-Tropism](https://huggingface.co/duttaprat/HViLM-Tropism)
-- **Benchmark:** [duttaprat/HVUE-v2](https://huggingface.co/datasets/duttaprat/HVUE-v2)
+- **Benchmark:** [duttaprat/HVUE](https://huggingface.co/datasets/duttaprat/HVUE)
 - **Code:** [github.com/duttaprat/HViLM](https://github.com/duttaprat/HViLM)
 - **Collections:** [duttaprat's Hugging Face Collections](https://huggingface.co/duttaprat/collections)