Refine model card

README.md

@@ -21,7 +21,7 @@ Metric with Feedback Signals*. The base model is
 [`google/gemma-3-4b-it`](https://huggingface.co/google/gemma-3-4b-it); base-model
 weights are not included here.
 
-This checkpoint corresponds to the paper's E2E setting and was trained on the joint WebNLG--E2E synthetic training set; the short repository name does not mean E2E-only training.
+This E2E checkpoint was trained on the joint WebNLG--E2E synthetic training set.
 
 ## Intended use
 
@@ -31,9 +31,8 @@ terminology, `missing` identifies an input unit omitted from the text; `extra`
 identifies text content unsupported by the input; and `incorrect` identifies an
 input unit realised with incorrect information.
 
-The expected prompt and four frozen regression examples are provided in
-`smoke_test.json`. This model is an evaluation component, not a general-purpose
-fact checker, and has not been validated outside data--text alignment settings.
+The expected prompt and four example inputs and outputs are provided in
+`smoke_test.json`.
 
 ## Prompt format
 
@@ -45,10 +44,9 @@ TRIPLES:
 Output as markdown table with Type and Triple columns.
 ```
 
-The canonical paper implementation uses **ms-swift PtEngine**. Transformers and
-vLLM are provided as portable alternatives. Different libraries, versions, and
-sampling implementations can produce small output differences; compare parsed
-error units rather than requiring byte-identical text.
+The canonical implementation uses **ms-swift PtEngine**. Transformers and vLLM
+examples are also provided. Outputs may vary slightly across runtimes; evaluation
+uses the parsed error units.
 
 ## ms-swift PtEngine
 
@@ -120,9 +118,7 @@ print(processor.decode(generated, skip_special_tokens=True))
 
 ## vLLM
 
-The reference vLLM environment uses NVIDIA H100 hardware and the pinned versions
-listed in `requirements.txt`. Other recent CUDA GPUs may also work, but are treated
-as best-effort environments and should be recorded in the smoke-test report.
+Install the versions listed in `requirements.txt` before running this example.
 
 ```python
 from huggingface_hub import snapshot_download
@@ -152,16 +148,10 @@ print(outputs[0].outputs[0].text)
 
 ## Reproducibility
 
-- Adapter SHA-256 and sanitized training hyperparameters: `training_manifest.json`
-- Frozen inputs and reference outputs: `smoke_test.json`
-- Paper: [https://openreview.net/forum?id=t1037gQHuf](https://openreview.net/forum?id=t1037gQHuf)
+- Training configuration: `training_manifest.json`
+- Example inputs and outputs: `smoke_test.json`
 - Code: [https://github.com/guihuzhang/xqdt](https://github.com/guihuzhang/xqdt)
 
-The original training run did not freeze a public Hugging Face revision for every
-base model. The standard base-model ID above replaces the machine-local cache path
-stored by the training framework. Users must comply with the corresponding base
-model's access terms and license.
-
 ## Citation
 
 ```bibtex