polish model card + embed training curve

README.md

@@ -2,58 +2,141 @@
 license: apache-2.0
 language:
 - en
+library_name: transformers
 pipeline_tag: summarization
+base_model: google/long-t5-tglobal-base
 tags:
 - longt5
 - summarization
 - meeting-summarization
-- qmsum
+- long-context
 metrics:
 - rouge
 - bertscore
+model-index:
+- name: longt5-meeting-summarization
+  results:
+  - task:
+      type: summarization
+      name: Meeting Summarization
+    metrics:
+    - type: rouge
+      name: ROUGE-1
+      value: 0.4632
+    - type: rouge
+      name: ROUGE-2
+      value: 0.2446
+    - type: rouge
+      name: ROUGE-L
+      value: 0.3970
+    - type: bertscore
+      name: BERTScore-F1
+      value: 0.8573
 ---
 
-# LongT5 — Meeting Summarization (prod)
+# LongT5 — Meeting Summarization
 
-Fine-tuned LongT5 for long-context meeting/transcript summarization.
+A [LongT5](https://huggingface.co/google/long-t5-tglobal-base) (`tglobal-base`) model fine-tuned to summarize long meeting transcripts into concise abstractive summaries. LongT5's transient-global attention handles very long inputs (up to **16,384 tokens** here), making it well suited for full-meeting transcripts that overflow a standard 512/1024-token encoder.
 
-## Eval results
+- **Base model:** `google/long-t5-tglobal-base`
+- **Task:** Abstractive summarization (English meetings/transcripts)
+- **Max input / output:** 16,384 / 512 tokens
+- **License:** Apache-2.0
 
-Held-out test set (`test_metrics.json`):
+![Validation ROUGE per epoch](training_curve.png)
 
-| metric   | score  |
-|----------|--------|
-| ROUGE-1  | 0.4632 |
-| ROUGE-2  | 0.2446 |
-| ROUGE-L  | 0.3970 |
-| test_loss| 2.8342 |
-
-Comparison vs previous production model (`full_comparison.json`) — this model (`new_model`) wins on every metric:
-
-| metric       | old    | **new (this)** |
-|--------------|--------|----------------|
-| ROUGE-1      | 0.4645 | **0.4687**     |
-| ROUGE-2      | 0.2454 | **0.2500**     |
-| ROUGE-L      | 0.3991 | **0.4032**     |
-| BERTScore-F  | 0.8567 | **0.8573**     |
-
-## Usage
+## Quick start
 
 ```python
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
 
-repo = "REPO_ID"  # e.g. your-username/longt5-meeting-summarization
+repo = "tuanhqv123/longt5-meeting-summarization"
 tok = AutoTokenizer.from_pretrained(repo)
 model = AutoModelForSeq2SeqLM.from_pretrained(repo)
 
-text = "..."  # meeting transcript
-inputs = tok(text, return_tensors="pt", truncation=True, max_length=8192)
-ids = model.generate(**inputs, max_new_tokens=256, num_beams=4)
-print(tok.decode(ids[0], skip_special_tokens=True))
+transcript = """..."""  # your meeting transcript
+inputs = tok(transcript, return_tensors="pt", truncation=True, max_length=16384)
+summary_ids = model.generate(
+    **inputs,
+    max_new_tokens=512,
+    num_beams=4,
+    no_repeat_ngram_size=3,
+    length_penalty=1.0,
+)
+print(tok.decode(summary_ids[0], skip_special_tokens=True))
 ```
 
-Or download just the weights:
+Download just the weights:
 
 ```bash
-hf download REPO_ID --local-dir ./best_model
+hf download tuanhqv123/longt5-meeting-summarization --local-dir ./best_model
+```
+
+## Evaluation
+
+Held-out **test** set (342 examples):
+
+| Metric       | Score  |
+|--------------|--------|
+| ROUGE-1      | 0.4632 |
+| ROUGE-2      | 0.2446 |
+| ROUGE-L      | 0.3970 |
+| BERTScore-F1 | 0.8573 |
+| Test loss    | 2.8342 |
+
+This checkpoint replaces a previous production model and **improves on every metric**:
+
+| Metric       | Previous | **This model** |
+|--------------|----------|----------------|
+| ROUGE-1      | 0.4645   | **0.4687**     |
+| ROUGE-2      | 0.2454   | **0.2500**     |
+| ROUGE-L      | 0.3991   | **0.4032**     |
+| BERTScore-F1 | 0.8567   | **0.8573**     |
+
+<sub>(comparison run on the same evaluation set with identical generation settings)</sub>
+
+## Training
+
+| Setting              | Value                          |
+|----------------------|--------------------------------|
+| Base model           | `google/long-t5-tglobal-base`  |
+| Epochs               | 10 (`load_best_model_at_end`, best by ROUGE-L) |
+| Batch size           | 1                              |
+| Learning rate        | 1e-5                           |
+| Warmup steps         | 100                            |
+| Max grad norm        | 1.0                            |
+| Label smoothing      | 0.1                            |
+| Precision            | BF16                           |
+| Max input / output   | 16,384 / 512 tokens            |
+| Data split           | 3,757 train / 427 val / 342 test |
+| Hardware             | 1× NVIDIA RTX 4090             |
+| Training time        | ~3.7 h                         |
+
+### Validation ROUGE per epoch
+
+| Epoch | ROUGE-1 | ROUGE-2 | ROUGE-L |
+|-------|---------|---------|---------|
+| 1  | 0.3660 | 0.1708 | 0.3128 |
+| 3  | 0.4375 | 0.2276 | 0.3764 |
+| 5  | 0.4618 | 0.2444 | 0.3988 |
+| 7  | 0.4706 | 0.2492 | 0.4036 |
+| 9  | 0.4737 | 0.2530 | 0.4068 |
+| **10** | **0.4744** | **0.2545** | **0.4091** |
+
+## Intended use & limitations
+
+- **Intended:** abstractive summarization of English meeting transcripts / long conversational text.
+- **Limitations:** English-only; trained on meeting-style data so it may transfer poorly to other domains (legal, medical, code). Like all abstractive summarizers it can hallucinate — verify facts before relying on summaries. Inputs beyond 16,384 tokens are truncated.
+
+## Citation
+
+Built on LongT5:
+
+```bibtex
+@article{guo2021longt5,
+  title={LongT5: Efficient Text-To-Text Transformer for Long Sequences},
+  author={Guo, Mandy and Ainslie, Joshua and Uthus, David and Ontanon, Santiago and Ni, Jianmo and Sung, Yun-Hsuan and Yang, Yinfei},
+  journal={arXiv preprint arXiv:2112.07916},
+  year={2021}
+}
 ```