Commit
•
98c13b3
1
Parent(s):
8df05d2
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,86 @@
|
|
1 |
---
|
2 |
license: cc-by-nc-4.0
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: cc-by-nc-4.0
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
widget:
|
6 |
+
- text: |
|
7 |
+
Who are you?
|
8 |
+
example_title: "Pirate Introduction"
|
9 |
---
|
10 |
+
|
11 |
+
![Llama3-Pirate-Talk-8b-v0.1.png](https://huggingface.co/phanerozoic/Llama3-Pirate-Talk-8b-v0.1/resolve/main/Llama3-Pirate-Talk-8b-v0.1.jpg)
|
12 |
+
|
13 |
+
# Llama3-Pirate-Talk-8b-v0.1
|
14 |
+
|
15 |
+
Llama3-Pirate-Talk-8b-v0.1, developed by phanerozoic, is the first pirate-themed fine-tune of the Llama3 8b model. This version is crafted to generate authentic pirate-themed content, seamlessly blending historical accuracy with modern fictional representations of pirate speech.
|
16 |
+
|
17 |
+
## Developed by:
|
18 |
+
- phanerozoic
|
19 |
+
|
20 |
+
## License:
|
21 |
+
- cc-by-nc-4.0
|
22 |
+
|
23 |
+
## Finetuned from:
|
24 |
+
- Llama-3-8B
|
25 |
+
|
26 |
+
## Version Control:
|
27 |
+
- Initial release of Llama3-Pirate-Talk-8b-v0.1, marking a new frontier in thematic language model applications.
|
28 |
+
|
29 |
+
## Model Overview:
|
30 |
+
Llama3-Pirate-Talk-8b-v0.1 excels at generating engaging and character-rich pirate dialogue, ideal for entertainment, gaming, and narrative projects. It is designed to perform well in both automated customer interaction platforms and interactive entertainment settings.
|
31 |
+
|
32 |
+
## Performance:
|
33 |
+
The model shows a robust capacity to maintain pirate dialect consistently, adding thematic depth to interactions. While it thrives in generating thematic content, it is less suited for tasks requiring precise technical responses.
|
34 |
+
|
35 |
+
## Direct Use:
|
36 |
+
Optimized for generating content in themed environments, particularly where engagement and character speech are valued over factual accuracy.
|
37 |
+
|
38 |
+
## Training Data:
|
39 |
+
The model was fine-tuned on an abstracted version of "Moby Dick," restructured to enhance pirate vernacular and themes, ensuring rich and varied linguistic inputs.
|
40 |
+
|
41 |
+
## Custom Stopping Strings:
|
42 |
+
To enhance output quality and thematic consistency, custom stopping strings include:
|
43 |
+
- "}}\n\n\n{{"
|
44 |
+
- "\n\n\n"
|
45 |
+
- "\n\nYou:"
|
46 |
+
- "You:"
|
47 |
+
- "\n\n"
|
48 |
+
- "\nYou:"
|
49 |
+
- "\n"
|
50 |
+
|
51 |
+
## Training Hyperparameters and Fine-Tuning Details:
|
52 |
+
\`\`\`yaml
|
53 |
+
micro_batch_size: 1
|
54 |
+
batch_size: 0
|
55 |
+
epochs: 1
|
56 |
+
learning_rate: "2e-5"
|
57 |
+
lr_scheduler_type: "linear"
|
58 |
+
lora_rank: 8
|
59 |
+
lora_alpha: 16
|
60 |
+
lora_dropout: 0.05
|
61 |
+
cutoff_len: 256
|
62 |
+
warmup_steps: 8
|
63 |
+
optimizer: "adamw_torch"
|
64 |
+
grad_accumulation: 1
|
65 |
+
train_runtime: 1697.081 seconds
|
66 |
+
total_flos: 1.3663655883177984e+16
|
67 |
+
train_loss: 1.7511341453808817
|
68 |
+
\`\`\`
|
69 |
+
|
70 |
+
## Testing and Evaluation:
|
71 |
+
During the testing phase, we conducted a series of evaluations to compare Llama3-Pirate-Talk-8b-v0.1 against the base Llama3 model. These tests involved complex navigational and general knowledge questions designed to assess the model's ability to maintain its thematic integrity while responding accurately to technically demanding prompts. The model demonstrated a strong thematic presence with consistent use of pirate vernacular. However, it showed limitations in handling high-precision technical content, which is an expected trade-off given its thematic specialization. These insights have been instrumental in identifying areas for further model refinement.
|
72 |
+
|
73 |
+
## Limitations:
|
74 |
+
Llama3-Pirate-Talk-8b-v0.1 is specifically tailored for pirate-themed content. It may not perform well in non-themed or general language tasks, where neutrality and technical precision are required.
|
75 |
+
|
76 |
+
## Compute Infrastructure:
|
77 |
+
The model was efficiently trained on an RTX 6000 Ada GPU in about half an hour, demonstrating the effective use of resources in creating specialized language models.
|
78 |
+
|
79 |
+
## Results:
|
80 |
+
The model consistently delivers pirate-themed content with a high degree of linguistic coherence and thematic accuracy. However, the depth of responses can vary, suggesting further fine-tuning could enhance its capability to handle complex queries.
|
81 |
+
|
82 |
+
## Acknowledgments:
|
83 |
+
Special thanks to the developers of the base Llama-3 model at Meta, whose open-source architecture was instrumental in developing this thematic model.
|
84 |
+
|
85 |
+
## Summary:
|
86 |
+
Llama3-Pirate-Talk-8b-v0.1 stands out for its unique ability to enrich thematic applications with authentic and engaging pirate dialogue. While it excels in themed content creation, its specialized nature makes it less adaptable to general-purpose tasks, highlighting its role as a niche model in the realm of AI-driven text generation.
|