lewtun HF staff commited on
Commit
b1bcda6
1 Parent(s): a5fa5bd

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +11 -8
README.md CHANGED
@@ -1,6 +1,9 @@
1
  ---
2
  tags:
3
  - generated_from_trainer
 
 
 
4
  model-index:
5
  - name: starchat-beta
6
  results: []
@@ -10,9 +13,9 @@ license: bigcode-openrail-m
10
 
11
  <img src="https://huggingface.co/HuggingFaceH4/starchat-beta/resolve/main/model_logo.png" alt="StarChat Beta Logo" width="800" style="margin-left:'auto' margin-right:'auto' display:'block'"/>
12
 
13
- # Model Card for StarChat Beta
14
 
15
- StarChat is a series of language models that are trained to act as helpful coding assistants. StarChat Beta is the second model in the series, and is a fine-tuned version of [StarCoderPlus](https://huggingface.co/bigcode/starcoderplus) that was trained on an ["uncensored"](https://erichartford.com/uncensored-models) variant of the [`openassistant-guanaco` dataset](https://huggingface.co/datasets/timdettmers/openassistant-guanaco). We found that removing the in-built alignment of the OpenAssistant dataset boosted performance on the [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) and made the model more helpful at coding tasks. However, this means that model is likely to generate problematic text when prompted to do so and should only be used for educational and research purposes.
16
 
17
  ## Model Details
18
 
@@ -25,7 +28,7 @@ StarChat is a series of language models that are trained to act as helpful codin
25
  - **License:** BigCode Open RAIL-M v1
26
  - **Finetuned from model:** [bigcode/starcoderplus](https://huggingface.co/bigcode/starcoderplus)
27
 
28
- ### Model Sources [optional]
29
 
30
  <!-- Provide the basic links for the model. -->
31
 
@@ -45,6 +48,7 @@ from transformers import pipeline
45
 
46
  pipe = pipeline("text-generation", model="HuggingFaceH4/starchat-beta", torch_dtype=torch.bfloat16, device_map="auto")
47
 
 
48
  prompt_template = "<|system|>\n<|end|>\n<|user|>\n{query}<|end|>\n<|assistant|>"
49
  prompt = prompt_template.format(query="How do I sort a list in Python?")
50
  # We use a special <|end|> token with ID 49155 to denote ends of a turn
@@ -56,21 +60,20 @@ outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.2, top_
56
 
57
  <!-- This section is meant to convey both technical and sociotechnical limitations. -->
58
 
59
- StarChat Alpha has not been aligned to human preferences with techniques like RLHF or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so).
60
  Models trained primarily on code data will also have a more skewed demographic bias commensurate with the demographics of the GitHub community, for more on this see the [StarCoder dataset](https://huggingface.co/datasets/bigcode/starcoderdata) which is derived from The Stack.
61
 
62
-
63
  Since the base model was pretrained on a large corpus of code, it may produce code snippets that are syntactically valid but semantically incorrect.
64
  For example, it may produce code that does not compile or that produces incorrect results.
65
  It may also produce code that is vulnerable to security exploits.
66
  We have observed the model also has a tendency to produce false URLs which should be carefully inspected before clicking.
67
 
68
- StarChat Alpha was fine-tuned from the base model [StarCoder Base](https://huggingface.co/bigcode/starcoderbase), please refer to its model card's [Limitations Section](https://huggingface.co/bigcode/starcoderbase#limitations) for relevant information.
69
  In particular, the model was evaluated on some categories of gender biases, propensity for toxicity, and risk of suggesting code completions with known security flaws; these evaluations are reported in its [technical report](https://drive.google.com/file/d/1cN-b9GnWtHzQRoE7M7gAEyivY0kl4BYs/view).
70
 
71
  ## Training and evaluation data
72
 
73
- StarChat Beta is trained on an ["uncensored"](https://erichartford.com/uncensored-models) variant of the [`openassistant-guanaco` dataset](https://huggingface.co/datasets/timdettmers/openassistant-guanaco). We applied the same [recipe](https://huggingface.co/datasets/ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered/blob/main/wizardlm_clean.py) used to filter the ShareGPT datasets behind the [WizardLM](https://huggingface.co/datasets/ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered).
74
 
75
  ## Training procedure
76
 
@@ -112,7 +115,7 @@ The following hyperparameters were used during training:
112
 
113
  ## Citation
114
 
115
- <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
116
 
117
  **BibTeX:**
118
 
 
1
  ---
2
  tags:
3
  - generated_from_trainer
4
+ widget:
5
+ - text: "How can I write a Python function to generate the nth Fibonacci number?"
6
+ - text: "How do I get the current date using shell commands? Explain how it works."
7
  model-index:
8
  - name: starchat-beta
9
  results: []
 
13
 
14
  <img src="https://huggingface.co/HuggingFaceH4/starchat-beta/resolve/main/model_logo.png" alt="StarChat Beta Logo" width="800" style="margin-left:'auto' margin-right:'auto' display:'block'"/>
15
 
16
+ # Model Card for StarChat
17
 
18
+ StarChat is a series of language models that are trained to act as helpful coding assistants. StarChat is the second model in the series, and is a fine-tuned version of [StarCoderPlus](https://huggingface.co/bigcode/starcoderplus) that was trained on an ["uncensored"](https://erichartford.com/uncensored-models) variant of the [`openassistant-guanaco` dataset](https://huggingface.co/datasets/timdettmers/openassistant-guanaco). We found that removing the in-built alignment of the OpenAssistant dataset boosted performance on the [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard) and made the model more helpful at coding tasks. However, this means that model is likely to generate problematic text when prompted to do so and should only be used for educational and research purposes.
19
 
20
  ## Model Details
21
 
 
28
  - **License:** BigCode Open RAIL-M v1
29
  - **Finetuned from model:** [bigcode/starcoderplus](https://huggingface.co/bigcode/starcoderplus)
30
 
31
+ ### Model Sources
32
 
33
  <!-- Provide the basic links for the model. -->
34
 
 
48
 
49
  pipe = pipeline("text-generation", model="HuggingFaceH4/starchat-beta", torch_dtype=torch.bfloat16, device_map="auto")
50
 
51
+ # We use a variant of ChatML to format each message
52
  prompt_template = "<|system|>\n<|end|>\n<|user|>\n{query}<|end|>\n<|assistant|>"
53
  prompt = prompt_template.format(query="How do I sort a list in Python?")
54
  # We use a special <|end|> token with ID 49155 to denote ends of a turn
 
60
 
61
  <!-- This section is meant to convey both technical and sociotechnical limitations. -->
62
 
63
+ StarChat has not been aligned to human preferences with techniques like RLHF or deployed with in-the-loop filtering of responses like ChatGPT, so the model can produce problematic outputs (especially when prompted to do so).
64
  Models trained primarily on code data will also have a more skewed demographic bias commensurate with the demographics of the GitHub community, for more on this see the [StarCoder dataset](https://huggingface.co/datasets/bigcode/starcoderdata) which is derived from The Stack.
65
 
 
66
  Since the base model was pretrained on a large corpus of code, it may produce code snippets that are syntactically valid but semantically incorrect.
67
  For example, it may produce code that does not compile or that produces incorrect results.
68
  It may also produce code that is vulnerable to security exploits.
69
  We have observed the model also has a tendency to produce false URLs which should be carefully inspected before clicking.
70
 
71
+ StarChat was fine-tuned from the base model [StarCoderPlus](https://huggingface.co/bigcode/starcoderplus), please refer to its model card's [Limitations Section](https://huggingface.co/bigcode/starcoderplus#limitations) for relevant information.
72
  In particular, the model was evaluated on some categories of gender biases, propensity for toxicity, and risk of suggesting code completions with known security flaws; these evaluations are reported in its [technical report](https://drive.google.com/file/d/1cN-b9GnWtHzQRoE7M7gAEyivY0kl4BYs/view).
73
 
74
  ## Training and evaluation data
75
 
76
+ StarChat is trained on an ["uncensored"](https://erichartford.com/uncensored-models) variant of the [`openassistant-guanaco` dataset](https://huggingface.co/datasets/timdettmers/openassistant-guanaco). We applied the same [recipe](https://huggingface.co/datasets/ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered/blob/main/wizardlm_clean.py) used to filter the ShareGPT datasets behind the [WizardLM](https://huggingface.co/datasets/ehartford/WizardLM_alpaca_evol_instruct_70k_unfiltered).
77
 
78
  ## Training procedure
79
 
 
115
 
116
  ## Citation
117
 
118
+ Although there isn't a blog post or paper associated with StarChat-β, you can find details on the earlier version in the blog post below:
119
 
120
  **BibTeX:**
121