310 lines
11 KiB
Markdown
310 lines
11 KiB
Markdown
---
|
|
base_model: unsloth/Llama-3.2-1B
|
|
tags:
|
|
- text-generation-inference
|
|
- transformers
|
|
- unsloth
|
|
- llama
|
|
- trl
|
|
- sft
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
---
|
|
|
|
<html lang="en">
|
|
<head>
|
|
<meta charset="UTF-8">
|
|
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
|
</head>
|
|
<div class="container">
|
|
<h1>GRMR-V3-L1B</h1>
|
|
<p>GRMR-V3-L1B is a fine-tuned version of <a href="https://huggingface.co/unsloth/Llama-3.2-1B">unsloth/Llama-3.2-1B</a> specifically optimized for grammar correction tasks.</p>
|
|
<div class="important-note">
|
|
<p><strong>IMPORTANT:</strong> Please ensure you are using the following sampler settings for optimal results:</p>
|
|
<pre><code>temperature = 0.7
|
|
frequency_penalty = 0.0
|
|
presence_penalty = 0.0
|
|
min_p = 0.01
|
|
top_p = 0.95
|
|
top_k = 40</code></pre>
|
|
</div>
|
|
<h2>Model description</h2>
|
|
<p>GRMR-V3-L1B is a grammar correction model built on Meta's Llama 3.2 1B base model. It has been fine-tuned on a large dataset of grammar correction examples to help improve text quality by fixing grammatical errors, punctuation, spelling, and other language issues.</p>
|
|
<p>The model uses a specialized chat template that structures inputs as "text" and outputs as "corrected" to maintain a clear distinction between original and corrected content.</p>
|
|
<h3>Model Details</h3>
|
|
<div class="model-meta">
|
|
<p><strong>Developed by:</strong> qingy2024</p>
|
|
<p><strong>Finetuned from model:</strong> <a href="https://huggingface.co/unsloth/Llama-3.2-1B">unsloth/Llama-3.2-1B</a></p>
|
|
<p><strong>Model type:</strong> Llama 3.2</p>
|
|
<p><strong>Language(s):</strong> English</p>
|
|
<p><strong>License:</strong> apache-2.0</p>
|
|
<p><strong>Training Dataset:</strong> <a href="https://huggingface.co/datasets/qingy2024/grmr-v4-60k">qingy2024/grmr-v4-60k</a></p>
|
|
</div>
|
|
<hr>
|
|
<h2>Manual Testing Examples</h2>
|
|
<p>Here are a few examples of grammar corrections this model can handle:</p>
|
|
<table>
|
|
<thead>
|
|
<tr>
|
|
<th>Original Text</th>
|
|
<th>Corrected Text</th>
|
|
</tr>
|
|
</thead>
|
|
<tbody>
|
|
<tr>
|
|
<td>i dont know weather to bring a umbrella today</td>
|
|
<td>I don't know whether to bring an umbrella today.</td>
|
|
</tr>
|
|
<tr>
|
|
<td>she go to the store yesterday</td>
|
|
<td>She went to the store yesterday.</td>
|
|
</tr>
|
|
<tr>
|
|
<td>they is going to be late for the meeting</td>
|
|
<td>They are going to be late for the meeting.</td>
|
|
</tr>
|
|
<tr>
|
|
<td>the cat laying on the floor all day</td>
|
|
<td>The cat is laying on the floor all day.</td>
|
|
</tr>
|
|
</tbody>
|
|
</table>
|
|
<hr>
|
|
<h2>Training procedure</h2>
|
|
<p>The model was fine-tuned using full parameter fine-tuning (not LoRA) on the GRMR-V4-60K dataset. The training was optimized using the Unsloth framework for efficient training of LLMs.</p>
|
|
<h3>Training hyperparameters</h3>
|
|
<ul>
|
|
<li><strong>Batch size:</strong> 8</li>
|
|
<li><strong>Gradient accumulation steps:</strong> 2</li>
|
|
<li><strong>Learning rate:</strong> 5e-5</li>
|
|
<li><strong>Epochs:</strong> 1</li>
|
|
<li><strong>Optimizer:</strong> AdamW (8-bit)</li>
|
|
<li><strong>Weight decay:</strong> 0.01</li>
|
|
<li><strong>LR scheduler:</strong> Cosine</li>
|
|
<li><strong>Warmup steps:</strong> 180</li>
|
|
<li><strong>Max sequence length:</strong> 16,384</li>
|
|
<li><strong>Training precision:</strong> Mixed precision (BF16 where available, FP16 otherwise)</li>
|
|
</ul>
|
|
<h2>Intended uses & limitations</h2>
|
|
<p>This model is designed for grammar correction tasks. It can be used to:</p>
|
|
<ul>
|
|
<li>Fix grammatical errors in written text</li>
|
|
<li>Correct punctuation</li>
|
|
<li>Address spelling mistakes</li>
|
|
<li>Improve sentence structure and clarity</li>
|
|
</ul>
|
|
<h3>Limitations</h3>
|
|
<ul>
|
|
<li>The model may struggle with highly technical or domain-specific content</li>
|
|
<li>It may not fully understand context-dependent grammar rules in all cases</li>
|
|
<li>Performance may vary for non-standard English or text with multiple errors</li>
|
|
</ul>
|
|
<h2>How to use</h2>
|
|
<p><code>llama.cpp</code> and projects based on it should be able to run this model like any others.</p>
|
|
<p>For pure <code>transformers</code> code, you can refer here:</p>
|
|
<pre><code class="language-python">from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
# Load model and tokenizer
|
|
model_name = "qingy2024/GRMR-V3-L1B"
|
|
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
|
model = AutoModelForCausalLM.from_pretrained(model_name)
|
|
text_to_correct = "i am going to the store tommorow and buy some thing for dinner"
|
|
messages = [
|
|
{"role": "user", "content": text_to_correct}
|
|
]
|
|
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
|
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
|
outputs = model.generate(
|
|
inputs["input_ids"],
|
|
max_new_tokens=512,
|
|
temperature=0.1, # NOTE: For best results, use the recommended temperature of 0.7
|
|
do_sample=True
|
|
)
|
|
corrected_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
|
print(corrected_text)</code></pre>
|
|
<h3>Using with the Hugging Face pipeline</h3>
|
|
<pre><code class="language-python">from transformers import pipeline
|
|
pipe = pipeline(
|
|
"text-generation",
|
|
model="qingy2024/GRMR-V3-L1B",
|
|
torch_dtype="auto",
|
|
device_map="auto"
|
|
)
|
|
messages = [
|
|
{"role": "user", "content": "i dont know weather to bring a umbrella today"}
|
|
]
|
|
result = pipe(
|
|
messages,
|
|
max_new_tokens=100,
|
|
temperature=0.1, # NOTE: For best results, use the recommended temperature of 0.7
|
|
do_sample=True,
|
|
return_full_text=False
|
|
)[0]["generated_text"]
|
|
print(result)</code></pre>
|
|
<p><em>Note: The Python examples above use <code>temperature=0.1</code> for reproducibility in quick tests. For optimal grammar correction quality, please use the recommended sampler settings, especially <code>temperature=0.7</code>.</em></p>
|
|
<h2>Custom Chat Template</h2>
|
|
<p class="chat-template-info">The model uses a custom chat template with special formatting for grammar correction:</p>
|
|
<ul>
|
|
<li>User inputs are formatted with <code><|start_header_id|>text<|end_header_id|></code> headers</li>
|
|
<li>Model outputs are formatted with <code><|start_header_id|>corrected<|end_header_id|></code> headers</li>
|
|
<li>Messages are separated by <code><|eot_id|></code> tokens</li>
|
|
<li>The chat template should work without any extra tweaking in vLLM or <code>llama.cpp</code>.</li>
|
|
</ul>
|
|
<p>An example of the applied template structure might look like:</p>
|
|
<pre><code><|begin_of_text|><|start_header_id|>text<|end_header_id|>
|
|
|
|
i dont know weather to bring a umbrella today<|eot_id|><|start_header_id|>corrected<|end_header_id|>
|
|
|
|
I don't know whether to bring an umbrella today.<|eot_id|></code></pre>
|
|
<p>(When generating, you would only provide the input up to <code><|start_header_id|>corrected<|end_header_id|>\n\n</code>)</p>
|
|
<h2>Training Dataset</h2>
|
|
<p>The model was fine-tuned on the <a href="https://huggingface.co/datasets/qingy2024/grmr-v4-60k">qingy2024/grmr-v4-60k</a> dataset, which contains 60,000 examples of original text and their grammatically corrected versions.</p>
|
|
<h2>Bias, Risks, and Limitations</h2>
|
|
<ul>
|
|
<li>The model may reflect biases present in the training data.</li>
|
|
<li>It may not perform equally well across different writing styles or domains.</li>
|
|
<li>The model might occasionally introduce errors or change the meaning of text.</li>
|
|
<li>It focuses on grammatical correctness rather than stylistic improvements.</li>
|
|
</ul>
|
|
<h2>Contact</h2>
|
|
<p>For questions or issues related to the model, please reach out via Hugging Face or by creating an issue in the repository.</p>
|
|
</div>
|
|
<style>
|
|
body {
|
|
font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif, "Apple Color Emoji", "Segoe UI Emoji", "Segoe UI Symbol";
|
|
line-height: 1.6;
|
|
margin: 0;
|
|
padding: 0;
|
|
background-color: #f8f9fa;
|
|
color: #333;
|
|
}
|
|
.container {
|
|
max-width: 1200px;
|
|
margin: 10px auto;
|
|
padding: 25px;
|
|
background-color: #ffffff;
|
|
border-radius: 8px;
|
|
box-shadow: 0 4px 12px rgba(0, 0, 0, 0.08);
|
|
}
|
|
h1, h2, h3 {
|
|
color: #0056b3; /* Primary Blue */
|
|
margin-top: 1.5em;
|
|
margin-bottom: 0.7em;
|
|
}
|
|
h1 {
|
|
text-align: center;
|
|
font-size: 2.2em;
|
|
border-bottom: 2px solid #e0e0e0;
|
|
padding-bottom: 0.5em;
|
|
margin-top: 0;
|
|
}
|
|
h2 {
|
|
font-size: 1.8em;
|
|
border-bottom: 1px solid #e9ecef;
|
|
padding-bottom: 0.3em;
|
|
}
|
|
h3 {
|
|
font-size: 1.4em;
|
|
color: #007bff; /* Lighter Blue for sub-headings */
|
|
}
|
|
p, li {
|
|
font-size: 1em;
|
|
color: #555;
|
|
}
|
|
a {
|
|
color: #007bff;
|
|
text-decoration: none;
|
|
}
|
|
a:hover {
|
|
text-decoration: underline;
|
|
color: #0056b3;
|
|
}
|
|
.important-note {
|
|
background-color: #e7f3ff; /* Light blue background */
|
|
border-left: 5px solid #007bff; /* Blue accent border */
|
|
margin: 20px 0px;
|
|
border-radius: 5px;
|
|
}
|
|
.important-note strong {
|
|
color: #0056b3;
|
|
font-weight: 600;
|
|
}
|
|
.important-note {
|
|
background-color: #d0e8ff;
|
|
padding: 0.05em 1.0em;
|
|
border-radius: 3px;
|
|
font-size: 0.9em;
|
|
}
|
|
code {
|
|
padding: 0.1em 0.4em;
|
|
border-radius: 3px;
|
|
font-size: 0.9em;
|
|
}
|
|
table {
|
|
width: 100%;
|
|
border-collapse: collapse;
|
|
margin: 20px 0;
|
|
box-shadow: 0 2px 4px rgba(0,0,0,0.05);
|
|
}
|
|
th, td {
|
|
border: 1px solid #dee2e6;
|
|
padding: 10px 12px;
|
|
text-align: left;
|
|
vertical-align: top;
|
|
}
|
|
th {
|
|
background-color: #e9ecef; /* Light gray for headers */
|
|
font-weight: 600;
|
|
color: #212529;
|
|
}
|
|
td:first-child {
|
|
/* font-style: italic; */
|
|
color: #444;
|
|
}
|
|
pre {
|
|
background-color: #f1f3f5;
|
|
padding: 15px;
|
|
border-radius: 5px;
|
|
overflow-x: auto;
|
|
border: 1px solid #ced4da;
|
|
font-size: 0.9em;
|
|
}
|
|
code {
|
|
font-family: "SFMono-Regular", Consolas, "Liberation Mono", Menlo, Courier, monospace;
|
|
background-color: #e9ecef;
|
|
padding: 0.2em 0.4em;
|
|
border-radius: 3px;
|
|
font-size: 0.9em;
|
|
}
|
|
pre code {
|
|
background-color: transparent;
|
|
padding: 0;
|
|
border-radius: 0;
|
|
font-size: 1em;
|
|
}
|
|
ul {
|
|
padding-left: 20px;
|
|
}
|
|
li {
|
|
margin-bottom: 0.5em;
|
|
}
|
|
hr {
|
|
border: none;
|
|
border-top: 1px solid #e0e0e0;
|
|
margin: 30px 0;
|
|
}
|
|
.model-meta {
|
|
background-color: #f8f9fa;
|
|
padding: 15px;
|
|
border-radius: 5px;
|
|
margin-bottom: 20px;
|
|
border: 1px solid #e9ecef;
|
|
}
|
|
.model-meta p { margin-bottom: 0.5em; }
|
|
.model-meta strong { color: #333; }
|
|
/* Specific styling for chat template explanation */
|
|
.chat-template-info span {
|
|
font-weight: bold;
|
|
color: #0056b3;
|
|
}
|
|
</style> |