Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
base_model:
|
| 6 |
+
- microsoft/DialoGPT-medium
|
| 7 |
+
model_type: gpt2
|
| 8 |
+
tags:
|
| 9 |
+
- conversational
|
| 10 |
+
- fine-tuned
|
| 11 |
+
- dialogpt
|
| 12 |
+
- kimi-k2
|
| 13 |
+
pipeline_tag: text-generation
|
| 14 |
+
library_name: transformers
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# DialoGPT-medium-distill-Kimi-K2-Instruct
|
| 18 |
+
|
| 19 |
+
This model is a fine-tuned version of __**microsoft/DialoGPT-medium**__, specialized for a custom persona and critical knowledge injection. It has been trained to balance conversational flexibility with specific factual recall.
|
| 20 |
+
|
| 21 |
+
## Model Description
|
| 22 |
+
|
| 23 |
+
- **Model type:** Causal Language Model
|
| 24 |
+
- **Language(s):** English
|
| 25 |
+
- **Base Model:** DialoGPT-medium
|
| 26 |
+
- **Total Parameters:** 406.3M _(Post-fine-tune expanded state)_
|
| 27 |
+
|
| 28 |
+
## Intended Uses & Limitations
|
| 29 |
+
|
| 30 |
+
This model is designed for creative assistant tasks and casual conversation.
|
| 31 |
+
- **Direct Use:** Chatbots, creative storytelling, and persona-driven interactions.
|
| 32 |
+
- **Limitations:** Due to the small dataset size and "creative" training, the model may occasionally hallucinate or provide non-literal answers (e.g., creative definitions of common objects).
|
| 33 |
+
|
| 34 |
+
## Training Procedure
|
| 35 |
+
The model underwent a full fine-tune on a custom dataset consisting of critical facts and casual chat examples.
|
| 36 |
+
|
| 37 |
+
### Training Hyperparameters
|
| 38 |
+
|
| 39 |
+
- **Learning Rate:** 2e-5
|
| 40 |
+
- **Epochs:** 5
|
| 41 |
+
- **Batch Size:** 4 _(with gradient accumulation)_
|
| 42 |
+
- **Precision:** Mixed Precision _(FP16)_
|
| 43 |
+
- **Loss achieved:** 3.264037
|
| 44 |
+
|
| 45 |
+
## Weight Analysis
|
| 46 |
+
|
| 47 |
+
Post-training analysis showed a significant shift in the LM Head weights _(Absolute Shift: 4.4164)_, indicating a strong adaptation to the new conversational style while maintaining structural grammar stability in the transformer layers.
|
| 48 |
+
|
| 49 |
+
## How to Use
|
| 50 |
+
```py
|
| 51 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
| 52 |
+
|
| 53 |
+
tokenizer = AutoTokenizer.from_pretrained("Fu01978/DialoGPT-medium-distill-Kimi-K2-Instruct")
|
| 54 |
+
model = AutoModelForCausalLM.from_pretrained("Fu01978/DialoGPT-medium-distill-Kimi-K2-Instruct")
|
| 55 |
+
|
| 56 |
+
# For best results, use a temperature between 0.7 and 0.85
|
| 57 |
+
```
|