初始化项目，由ModelHub XC社区提供模型

Model: Arcoson/distilgpt2-arxiv-csai
Source: Original Platform

README.md

@@ -0,0 +1,154 @@
+---
+language:
+- en
+license: mit
+library_name: transformers
+tags:
+- text-generation
+- arxiv
+- computer-science
+- artificial-intelligence
+- research
+- gpt2
+- distilgpt2
+- fine-tuned
+datasets:
+- ccdv/arxiv-classification
+metrics:
+- perplexity
+base_model: distilgpt2
+pipeline_tag: text-generation
+model-index:
+- name: distilgpt2-arxiv-csai
+  results:
+  - task:
+      type: text-generation
+      name: Text Generation
+    dataset:
+      name: arXiv CS/AI Abstracts
+      type: ccdv/arxiv-classification
+    metrics:
+    - type: perplexity
+      value: 32.59
+      name: Perplexity
+---
+
+# DistilGPT-2 Fine-Tuned on arXiv CS/AI Abstracts
+
+A **DistilGPT-2** (82M parameters) model fine-tuned on computer science and artificial intelligence paper abstracts from arXiv. The model generates text in the style of academic CS/AI research abstracts.
+
+## Model Details
+
+| Property | Value |
+|---|---|
+| **Base Model** | [distilgpt2](https://huggingface.co/distilgpt2) |
+| **Parameters** | 81.9M total, 14.2M trainable (17.3%) |
+| **Fine-tuning Strategy** | Partial freeze — last 2 transformer blocks + LM head |
+| **Training Data** | [ccdv/arxiv-classification](https://huggingface.co/datasets/ccdv/arxiv-classification) (CS/AI subset) |
+| **Training Samples** | 2,000 |
+| **Max Sequence Length** | 128 tokens |
+| **License** | MIT |
+
+## Training Details
+
+### Dataset
+The model was fine-tuned on abstracts from the [ccdv/arxiv-classification](https://huggingface.co/datasets/ccdv/arxiv-classification) dataset, filtered for computer science and AI categories:
+
+- `cs.CV` (Computer Vision)
+- `cs.AI` (Artificial Intelligence)
+- `cs.SY` (Systems and Control)
+- `cs.CE` (Computational Engineering)
+- `cs.PL` (Programming Languages)
+- `cs.IT` (Information Theory)
+- `cs.DS` (Data Structures and Algorithms)
+- `cs.NE` (Neural and Evolutionary Computing)
+
+### Hyperparameters
+
+| Parameter | Value |
+|---|---|
+| Learning rate | 3e-4 |
+| Batch size | 8 |
+| Max steps | 100 |
+| Warmup steps | 10 |
+| Scheduler | Cosine |
+| Weight decay | 0.01 |
+| Frozen layers | Embedding + blocks 0-3 |
+| Trainable layers | Blocks 4-5 + LayerNorm + LM Head |
+
+### Metrics
+
+| Metric | Value |
+|---|---|
+| Training loss | 3.82 |
+| Eval loss | 3.48 |
+| **Perplexity** | **32.59** |
+| Training time | ~7 minutes (CPU) |
+
+## Usage
+
+```python
+from transformers import pipeline
+
+generator = pipeline("text-generation", model="Arcoson/distilgpt2-arxiv-csai")
+
+output = generator(
+    "arXiv Abstract: We propose a novel approach to",
+    max_new_tokens=100,
+    do_sample=True,
+    temperature=0.8,
+    top_p=0.9,
+)
+print(output[0]["generated_text"])
+```
+
+Or load directly:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("Arcoson/distilgpt2-arxiv-csai")
+model = AutoModelForCausalLM.from_pretrained("Arcoson/distilgpt2-arxiv-csai")
+
+inputs = tokenizer("arXiv Abstract: In this paper, we study the problem of", return_tensors="pt")
+outputs = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.8)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+## Sample Generations
+
+**Prompt:** "arXiv Abstract: We propose a novel approach to"
+
+**Fine-tuned output:**
+> We propose a novel approach to the problem of time-time in the computation model, i.e., and prediction models for (i) continuous intervals which are nonlinear because they cannot be solved with finite random variables or large number functions...
+
+**Base DistilGPT-2 output:**
+> We propose a novel approach to the question of whether we can solve this problem by using real-time neural networks in an MRI technique...
+
+The fine-tuned model produces more domain-appropriate language with CS/AI terminology compared to the base model.
+
+## Limitations
+
+- **Short context**: Trained on 128-token sequences, so it works best for generating short abstracts or completions.
+- **Limited training**: Only 100 steps on 2,000 samples — this is a proof-of-concept fine-tune. More training data and steps would improve quality.
+- **Domain narrow**: Outputs are biased toward CS/AI academic language. Not suitable for general-purpose text generation.
+- **No factual accuracy**: The model generates plausible-sounding but not necessarily factually correct research text.
+
+## Training Infrastructure
+
+- **Hardware**: CPU (2 vCPUs, 8GB RAM)
+- **Framework**: HuggingFace Transformers 5.5.0, PyTorch 2.11
+- **Strategy**: Partial layer freezing to enable CPU-feasible training
+
+## Citation
+
+If you use this model, please cite the base model and dataset:
+
+```bibtex
+@article{sanh2019distilbert,
+  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
+  author={Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas},
+  journal={arXiv preprint arXiv:1910.01108},
+  year={2019}
+}
+```