Update README.md

README.md

@@ -1,12 +1,16 @@
 ---
 library_name: transformers
-tags: []
+license: apache-2.0
+language:
+- en
+base_model:
+- meta-llama/Llama-3.1-8B-Instruct
 ---
 
 # Model Card for Model ID
 
 <!-- Provide a quick summary of what the model is/does. -->
-
+VeriFastScore is a factuality evaluation model designed for long-form LLM outputs. It jointly extracts and verifies factual claims in a single model pass, providing a faster alternative to pipeline-based evaluators like VeriScore.
 
 
 ## Model Details
@@ -14,24 +18,20 @@ tags: []
 ### Model Description
 
 <!-- Provide a longer summary of what this model is. -->
+This is a fine-tuned LLaMA 3.1 8B Instruct model trained to extract and verify factual claims in long-form text, given associated retrieved evidence. The model is designed to reduce inference latency and cost while maintaining high agreement with more expensive pipeline-based factuality metrics.
 
-This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
-
-- **Developed by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Model type:** [More Information Needed]
-- **Language(s) (NLP):** [More Information Needed]
-- **License:** [More Information Needed]
-- **Finetuned from model [optional]:** [More Information Needed]
+- **Developed by:** NGRAM at UMD, Lambda Labs
+- **Model type:** Factuality evaluation model (joint claim extraction and verification) (Causal LM)
+- **Language(s) (NLP):** English
+- **License:** Apache 2.0
+- **Finetuned from model:** meta-llama/Llama-3.1-8B-Instruct
 
-### Model Sources [optional]
+### Model Sources
 
 <!-- Provide the basic links for the model. -->
 
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
+- **Repository:** <a href="https://github.com/RishanthRajendhran/VeriFastScore">github.com/RishanthRajendhran/VeriFastScore</a>
+- **Paper:** <a href="https://arxiv.org/abs/2505.16973">arxiv.org/abs/2505.16973</a>
 
 ## Uses
 
@@ -40,65 +40,77 @@ This is the model card of a 🤗 transformers model that has been pushed on the
 ### Direct Use
 
 <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+The model takes as input a generated long-form response and a consolidated set of retrieved evidence sentences. It outputs a list of verifiable claims and corresponding factuality labels (Supported or Unsupported).
 
-[More Information Needed]
-
-### Downstream Use [optional]
+### Downstream Use
 
 <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
-
-[More Information Needed]
+Can be used to score factuality in evaluation pipelines (e.g., RLHF supervision), dataset filtering, or system-level benchmarking of LLM factuality.
 
 ### Out-of-Scope Use
 
 <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
-
-[More Information Needed]
+- Not intended for use without retrieved evidence.
 
 ## Bias, Risks, and Limitations
 
 <!-- This section is meant to convey both technical and sociotechnical limitations. -->
-
-[More Information Needed]
+The model inherits potential biases from its teacher supervision (VeriScore) and the base language model. It may underperform on ambiguous claims, noisy evidence, or non-English text.
 
 ### Recommendations
 
 <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
 
-Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+Use caution in high-stakes domains and supplement with human review if used for system-level feedback or alignment. Avoid use cases without explicit, relevant evidence input.
 
 ## How to Get Started with the Model
 
 Use the code below to get started with the model.
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+tokenizer = AutoTokenizer.from_pretrained("rishanthrajendhran/VeriFastScore")
+model = AutoModelForCausalLM.from_pretrained("rishanthrajendhran/VeriFastScore")
 
-[More Information Needed]
+prompt = "<your prompt with evidence and response>"
+inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
+outputs = model.generate(**inputs, max_new_tokens=512)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
 
 ## Training Details
 
 ### Training Data
 
 <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
-
-[More Information Needed]
+Synthetic (response, evidence, claim, label) examples generated via VeriScore applied to long-form prompts from datasets like Tulu3-Personas. See <a href="https://arxiv.org/abs/2505.16973" style="color:black;">paper</a> for more details.
 
 ### Training Procedure
 
 <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+Two-stage fine-tuning:
+
+- Stage 1: Supervision with claim-level evidence.
+- Stage 2: Supervision with a mixture of claim- and sentence-level evidence.
 
-#### Preprocessing [optional]
+#### Preprocessing
 
-[More Information Needed]
+In the original VeriFastScore pipeline, evidence is aggregated at the sentence level per response, tokenized, and paired with output claims using a structured prompt template. However, the \VeriFastScore model is agnostic to the provenance of the provided evidence.
 
 
 #### Training Hyperparameters
 
-- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+- **Training regime:** : bf16 mixed precision
+- **Optimizer**: AdamW
+- **Scheduler**: Cosine decay (optional placeholder)
+- **Batch size**: 8 (effective)
+- **Epochs**: 10 (5+5)
 
-#### Speeds, Sizes, Times [optional]
+#### Speeds, Sizes, Times
 
 <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
-
-[More Information Needed]
+- Training Time: ~24*4 GPU hours (roughly 2 sec per training instance)
+- Model Size: 8B parameters
 
 ## Evaluation
 
@@ -109,34 +121,32 @@ Use the code below to get started with the model.
 #### Testing Data
 
 <!-- This should link to a Dataset Card if possible. -->
-
-[More Information Needed]
-
-#### Factors
-
-<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
-
-[More Information Needed]
+- ~9k test instances using both claim-level and sentence-level evidence
+- Model rankings: 100 prompts from the Tulu3-Personas test set with responses from 12 LLMs
 
 #### Metrics
 
 <!-- These are the evaluation metrics being used, ideally with a description of why. -->
-
-[More Information Needed]
+- Claim-level accuracy, precision, recall (automatic judgements using GPT-4o-mini)
+- Pearson correlation with factuality scores from VeriScore
 
 ### Results
-
-[More Information Needed]
+- (Claim-level evidence) Pearson r with VeriScore: 0.86, p<0.001
+- (Sentence-level evidence) Pearson r with VeriScore: 0.80, p<0.001
+- Model rankings:
+  - System-level Pearson r: 0.94, p<0.001
+  - Speedup: 6.6× (9.9× if excluding retrieval)
+See paper for more details.
 
 #### Summary
+VeriFastScore delivers fast, interpretable factuality scores that closely track a strong multi-step baseline, while reducing cost and latency for large-scale evaluation.
 
 
 
-## Model Examination [optional]
+## Model Examination
 
 <!-- Relevant interpretability work for the model goes here -->
-
-[More Information Needed]
+Future work could explore explainability or rationale generation via mode-switching techniques or chain-of-thought prompting.
 
 ## Environmental Impact
 
@@ -144,56 +154,28 @@ Use the code below to get started with the model.
 
 Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
 
-- **Hardware Type:** [More Information Needed]
-- **Hours used:** [More Information Needed]
-- **Cloud Provider:** [More Information Needed]
-- **Compute Region:** [More Information Needed]
-- **Carbon Emitted:** [More Information Needed]
-
-## Technical Specifications [optional]
-
-### Model Architecture and Objective
-
-[More Information Needed]
-
-### Compute Infrastructure
-
-[More Information Needed]
-
-#### Hardware
-
-[More Information Needed]
-
-#### Software
-
-[More Information Needed]
+- **Hardware Type:** A100 (Training), GH200 (Evaluation, Testing)
+- **Hours used:** 96 (Training)
+- **Cloud Provider:** Lambda Labs
+- **Compute Region:** us-central1
+- **Carbon Emitted:** 10.37 (Training)
 
 ## Citation [optional]
 
-<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
-
 **BibTeX:**
 
-[More Information Needed]
-
-**APA:**
-
-[More Information Needed]
-
-## Glossary [optional]
-
-<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
-
-[More Information Needed]
-
-## More Information [optional]
-
-[More Information Needed]
-
-## Model Card Authors [optional]
-
-[More Information Needed]
+<pre>
+  @misc{rajendhran2025verifastscorespeedinglongformfactuality,
+      title={VeriFastScore: Speeding up long-form factuality evaluation}, 
+      author={Rishanth Rajendhran and Amir Zadeh and Matthew Sarte and Chuan Li and Mohit Iyyer},
+      year={2025},
+      eprint={2505.16973},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL},
+      url={https://arxiv.org/abs/2505.16973}, 
+}
+</pre>
 
 ## Model Card Contact
 
-[More Information Needed]
+rishanth@umd.edu