Upload folder using huggingface_hub
Browse files- README.md +597 -0
- adapter_index.json +151 -0
- chat_template.jinja +185 -0
- compose_report.json +0 -0
- config.json +233 -0
- generation_config.json +10 -0
- io_configs/answerability/io.yaml +21 -0
- io_configs/citations/io.yaml +98 -0
- io_configs/context-attribution/io.yaml +100 -0
- io_configs/context_relevance/io.yaml +25 -0
- io_configs/factuality-correction/io.yaml +23 -0
- io_configs/factuality-detection/io.yaml +24 -0
- io_configs/guardian-core/io.yaml +32 -0
- io_configs/hallucination_detection/io.yaml +73 -0
- io_configs/policy-guardrails/io.yaml +24 -0
- io_configs/query_clarification/io.yaml +24 -0
- io_configs/query_rewrite/io.yaml +24 -0
- io_configs/requirement-check/io.yaml +32 -0
- io_configs/uncertainty/io.yaml +42 -0
- merges.txt +0 -0
- model.safetensors +3 -0
- special_tokens_map.json +30 -0
- tokenizer.json +0 -0
- tokenizer_config.json +30 -0
- vocab.json +0 -0
README.md
ADDED
|
@@ -0,0 +1,597 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: transformers
|
| 4 |
+
tags:
|
| 5 |
+
- language
|
| 6 |
+
- granite-4.1
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
[](https://mot.isitopen.ai/model/1160)
|
| 10 |
+
|
| 11 |
+
# Granite-4.1-3B
|
| 12 |
+
|
| 13 |
+
<!-- 📣 **Update [10-07-2025]:** Added a *default system prompt* to the chat template to guide the model towards more *professional, accurate, and safe* responses. -->
|
| 14 |
+
|
| 15 |
+
**Model Summary:**
|
| 16 |
+
Granite-4.1-3B is a 3B parameter long-context instruct model finetuned from *Granite-4.1-3B-Base* using a combination of open source instruction datasets with permissive license and internally collected synthetic datasets. Granite 4.1 models have gone through an improved post-training pipeline, including supervised finetuning and reinforcement learning alignment, resulting in enhanced tool calling, instruction following, and chat capabilities.
|
| 17 |
+
|
| 18 |
+
- **Developers:** Granite Team, IBM
|
| 19 |
+
- **HF Collection:** [Granite 4.1 Language Models HF Collection](https://huggingface.co/collections/ibm-granite/granite-41-language-models)
|
| 20 |
+
- **Technical Blog:** [Granite-4.1 Blog](https://huggingface.co/blog/ibm-granite/granite-4-1)
|
| 21 |
+
- **GitHub Repository:** [ibm-granite/granite-4.1-language-models](https://github.com/ibm-granite/granite-4.1-language-models)
|
| 22 |
+
- **Website**: [Granite Docs](https://www.ibm.com/granite/docs/)
|
| 23 |
+
- **Release Date**: April 29th, 2026
|
| 24 |
+
- **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)
|
| 25 |
+
|
| 26 |
+
**Supported Languages:**
|
| 27 |
+
English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, and Chinese. Users may finetune Granite 4.1 models for languages beyond these languages.
|
| 28 |
+
|
| 29 |
+
**Intended use:**
|
| 30 |
+
The model is designed to follow general instructions and can serve as the foundation for AI assistants across diverse domains, including business applications, as well as for LLM agents equipped with tool-use capabilities.
|
| 31 |
+
|
| 32 |
+
*Capabilities*
|
| 33 |
+
* Summarization
|
| 34 |
+
* Text classification
|
| 35 |
+
* Text extraction
|
| 36 |
+
* Question-answering
|
| 37 |
+
* Retrieval Augmented Generation (RAG)
|
| 38 |
+
* Code related tasks
|
| 39 |
+
* Function-calling tasks
|
| 40 |
+
* Multilingual dialog use cases
|
| 41 |
+
* Fill-In-the-Middle (FIM) code completions
|
| 42 |
+
|
| 43 |
+
<!-- <todo>Need to test the examples. (especially the tool calling and RAG ones)</todo>
|
| 44 |
+
-->
|
| 45 |
+
|
| 46 |
+
**Generation:**
|
| 47 |
+
This is a simple example of how to use Granite-4.1-3B model.
|
| 48 |
+
|
| 49 |
+
Install the following libraries:
|
| 50 |
+
|
| 51 |
+
```shell
|
| 52 |
+
pip install torch torchvision torchaudio
|
| 53 |
+
pip install accelerate
|
| 54 |
+
pip install transformers
|
| 55 |
+
```
|
| 56 |
+
Then, copy the snippet from the section that is relevant for your use case.
|
| 57 |
+
|
| 58 |
+
```python
|
| 59 |
+
import torch
|
| 60 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 61 |
+
|
| 62 |
+
device = "cuda"
|
| 63 |
+
model_path = "ibm-granite/granite-4.1-3b"
|
| 64 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
| 65 |
+
# drop device_map if running on CPU
|
| 66 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device)
|
| 67 |
+
model.eval()
|
| 68 |
+
# change input text as desired
|
| 69 |
+
chat = [
|
| 70 |
+
{ "role": "user", "content": "Please list one IBM Research laboratory located in the United States. You should only output its name and location." },
|
| 71 |
+
]
|
| 72 |
+
chat = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
|
| 73 |
+
# tokenize the text
|
| 74 |
+
input_tokens = tokenizer(chat, return_tensors="pt").to(device)
|
| 75 |
+
# generate output tokens
|
| 76 |
+
output = model.generate(**input_tokens,
|
| 77 |
+
max_new_tokens=100)
|
| 78 |
+
# decode output tokens into text
|
| 79 |
+
output = tokenizer.batch_decode(output)
|
| 80 |
+
# print output
|
| 81 |
+
print(output[0])
|
| 82 |
+
```
|
| 83 |
+
|
| 84 |
+
Expected output:
|
| 85 |
+
```shell
|
| 86 |
+
<|start_of_role|>user<|end_of_role|>Please list one IBM Research laboratory located in the United States. You should only output its name and location.<|end_of_text|>
|
| 87 |
+
<|start_of_role|>assistant<|end_of_role|>Almaden Research Center, San Jose, California<|end_of_text|>
|
| 88 |
+
```
|
| 89 |
+
<!-- 📣 **Update [2025-10-07]:** Added a *default system prompt* to the chat template to guide the model towards more *professional, accurate, and safe* responses. -->
|
| 90 |
+
|
| 91 |
+
**Tool-calling:**
|
| 92 |
+
Granite-4.1-3B comes with enhanced tool calling capabilities, enabling seamless integration with external functions and APIs. To define a list of tools please follow OpenAI's function [definition schema](https://platform.openai.com/docs/guides/function-calling?api-mode=responses#defining-functions).
|
| 93 |
+
|
| 94 |
+
This is an example of how to use Granite-4.1-3B model tool-calling ability:
|
| 95 |
+
|
| 96 |
+
```python
|
| 97 |
+
import torch
|
| 98 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 99 |
+
|
| 100 |
+
device = "cuda"
|
| 101 |
+
model_path = "ibm-granite/granite-4.1-3b"
|
| 102 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
| 103 |
+
# drop device_map if running on CPU
|
| 104 |
+
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device)
|
| 105 |
+
model.eval()
|
| 106 |
+
|
| 107 |
+
tools = [
|
| 108 |
+
{
|
| 109 |
+
"type": "function",
|
| 110 |
+
"function": {
|
| 111 |
+
"name": "get_current_weather",
|
| 112 |
+
"description": "Get the current weather for a specified city.",
|
| 113 |
+
"parameters": {
|
| 114 |
+
"type": "object",
|
| 115 |
+
"properties": {
|
| 116 |
+
"city": {
|
| 117 |
+
"type": "string",
|
| 118 |
+
"description": "Name of the city"
|
| 119 |
+
}
|
| 120 |
+
},
|
| 121 |
+
"required": ["city"]
|
| 122 |
+
}
|
| 123 |
+
}
|
| 124 |
+
}
|
| 125 |
+
]
|
| 126 |
+
|
| 127 |
+
# change input text as desired
|
| 128 |
+
chat = [
|
| 129 |
+
{ "role": "user", "content": "What's the weather like in Boston right now?" },
|
| 130 |
+
]
|
| 131 |
+
chat = tokenizer.apply_chat_template(chat, \
|
| 132 |
+
tokenize=False, \
|
| 133 |
+
tools=tools, \
|
| 134 |
+
add_generation_prompt=True)
|
| 135 |
+
# tokenize the text
|
| 136 |
+
input_tokens = tokenizer(chat, return_tensors="pt").to(device)
|
| 137 |
+
# generate output tokens
|
| 138 |
+
output = model.generate(**input_tokens,
|
| 139 |
+
max_new_tokens=100)
|
| 140 |
+
# decode output tokens into text
|
| 141 |
+
output = tokenizer.batch_decode(output)
|
| 142 |
+
# print output
|
| 143 |
+
print(output[0])
|
| 144 |
+
```
|
| 145 |
+
|
| 146 |
+
Expected output:
|
| 147 |
+
```shell
|
| 148 |
+
<|start_of_role|>system<|end_of_role|>You are a helpful assistant with access to the following tools. You may call one or more tools to assist with the user query.
|
| 149 |
+
|
| 150 |
+
You are provided with function signatures within <tools></tools> XML tags:
|
| 151 |
+
<tools>
|
| 152 |
+
{"type": "function", "function": {"name": "get_current_weather", "description": "Get the current weather for a specified city.", "parameters": {"type": "object", "properties": {"city": {"type": "string", "description": "Name of the city"}}, "required": ["city"]}}}
|
| 153 |
+
</tools>
|
| 154 |
+
|
| 155 |
+
For each tool call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
|
| 156 |
+
<tool_call>
|
| 157 |
+
{"name": <function-name>, "arguments": <args-json-object>}
|
| 158 |
+
</tool_call>. If a tool does not exist in the provided list of tools, notify the user that you do not have the ability to fulfill the request.<|end_of_text|>
|
| 159 |
+
<|start_of_role|>user<|end_of_role|>What's the weather like in Boston right now?<|end_of_text|>
|
| 160 |
+
<|start_of_role|>assistant<|end_of_role|><tool_call>
|
| 161 |
+
{"name": "get_current_weather", "arguments": {"city": "Boston"}}
|
| 162 |
+
</tool_call><|end_of_text|>
|
| 163 |
+
```
|
| 164 |
+
|
| 165 |
+
<!-- **Retrieval Augmented Generation:**
|
| 166 |
+
*Coming soon* -->
|
| 167 |
+
|
| 168 |
+
**Evaluation Results:**
|
| 169 |
+
|
| 170 |
+
<table>
|
| 171 |
+
<!-- <caption><b> All Results</b></caption> -->
|
| 172 |
+
<thead>
|
| 173 |
+
<tr>
|
| 174 |
+
<th style="text-align:left; background-color: #001d6c; color: white;">Benchmarks</th>
|
| 175 |
+
<th style="text-align:left; background-color: #001d6c; color: white;">Metric</th>
|
| 176 |
+
<th style="text-align:center; background-color: #001d6c; color: white;">3B Dense</th>
|
| 177 |
+
<th style="text-align:center; background-color: #001d6c; color: white;">8B Dense</th>
|
| 178 |
+
<th style="text-align:center; background-color: #001d6c; color: white;">30B Dense</th>
|
| 179 |
+
</tr>
|
| 180 |
+
</thead>
|
| 181 |
+
<tbody>
|
| 182 |
+
<tr>
|
| 183 |
+
<td colspan="5" style="text-align:center; background-color: #FFFFFF; color: #2D2D2D; font-style:italic;">
|
| 184 |
+
General Tasks
|
| 185 |
+
</td>
|
| 186 |
+
</tr>
|
| 187 |
+
<tr>
|
| 188 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MMLU</td>
|
| 189 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">5-shot</td>
|
| 190 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">67.02</td>
|
| 191 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">73.84</td>
|
| 192 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">80.16</td>
|
| 193 |
+
</tr>
|
| 194 |
+
<tr>
|
| 195 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MMLU-Pro</td>
|
| 196 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">5-shot, CoT</td>
|
| 197 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">49.83</td>
|
| 198 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">55.99</td>
|
| 199 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">64.09</td>
|
| 200 |
+
</tr>
|
| 201 |
+
<tr>
|
| 202 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">BBH</td>
|
| 203 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">3-shot, CoT</td>
|
| 204 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">75.83</td>
|
| 205 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">80.51</td>
|
| 206 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">83.74</td>
|
| 207 |
+
</tr>
|
| 208 |
+
<tr>
|
| 209 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">AGI EVAL</td>
|
| 210 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">0-shot, CoT</td>
|
| 211 |
+
<td style="text-align:right; background-color:#DAE8FF; color: #2D2D2D;">65.16</td>
|
| 212 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">72.43</td>
|
| 213 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">77.80</td>
|
| 214 |
+
</tr>
|
| 215 |
+
<tr>
|
| 216 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">GPQA</td>
|
| 217 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">0-shot, CoT</td>
|
| 218 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">31.70</td>
|
| 219 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">41.96</td>
|
| 220 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">45.76</td>
|
| 221 |
+
</tr>
|
| 222 |
+
<tr>
|
| 223 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">SimpleQA</td>
|
| 224 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 225 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">3.68</td>
|
| 226 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">4.82</td>
|
| 227 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">6.81</td>
|
| 228 |
+
</tr>
|
| 229 |
+
<tr>
|
| 230 |
+
<td colspan="5" style="text-align:center; background-color: #FFFFFF; color: #2D2D2D; font-style:italic;">
|
| 231 |
+
Alignment Tasks
|
| 232 |
+
</td>
|
| 233 |
+
</tr>
|
| 234 |
+
<tr>
|
| 235 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">AlpacaEval 2.0</td>
|
| 236 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 237 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">38.57</td>
|
| 238 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">50.08</td>
|
| 239 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">56.16</td>
|
| 240 |
+
</tr>
|
| 241 |
+
<tr>
|
| 242 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">IFEval Avg</td>
|
| 243 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 244 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">82.30</td>
|
| 245 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">87.06</td>
|
| 246 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">89.65</td>
|
| 247 |
+
</tr>
|
| 248 |
+
<tr>
|
| 249 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">ArenaHard</td>
|
| 250 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 251 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">37.80</td>
|
| 252 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">68.98</td>
|
| 253 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">71.02</td>
|
| 254 |
+
</tr>
|
| 255 |
+
<tr>
|
| 256 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MTBench Avg</td>
|
| 257 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 258 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">7.57</td>
|
| 259 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">8.61</td>
|
| 260 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">8.61</td>
|
| 261 |
+
</tr>
|
| 262 |
+
<tr>
|
| 263 |
+
<td colspan="5" style="text-align:center; background-color: #FFFFFF; color: #2D2D2D; font-style:italic;">
|
| 264 |
+
Math Tasks
|
| 265 |
+
</td>
|
| 266 |
+
</tr>
|
| 267 |
+
<tr>
|
| 268 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">GSM8K</td>
|
| 269 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">8-shot</td>
|
| 270 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">86.88</td>
|
| 271 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">92.49</td>
|
| 272 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">94.16</td>
|
| 273 |
+
</tr>
|
| 274 |
+
<tr>
|
| 275 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">GSM Symbolic</td>
|
| 276 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">8-shot</td>
|
| 277 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">81.32</td>
|
| 278 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">83.70</td>
|
| 279 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">75.70</td>
|
| 280 |
+
</tr>
|
| 281 |
+
<tr>
|
| 282 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">Minerva Math</td>
|
| 283 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">0-shot, CoT</td>
|
| 284 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">67.94</td>
|
| 285 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">80.10</td>
|
| 286 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">81.32</td>
|
| 287 |
+
</tr>
|
| 288 |
+
<tr>
|
| 289 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">DeepMind Math</td>
|
| 290 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">0-shot, CoT</td>
|
| 291 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">64.64</td>
|
| 292 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">80.07</td>
|
| 293 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">81.93</td>
|
| 294 |
+
</tr>
|
| 295 |
+
<tr>
|
| 296 |
+
<td colspan="5" style="text-align:center; background-color: #FFFFFF; color: #2D2D2D; font-style:italic;">
|
| 297 |
+
Code Tasks
|
| 298 |
+
</td>
|
| 299 |
+
</tr>
|
| 300 |
+
<tr>
|
| 301 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">HumanEval</td>
|
| 302 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">pass@1</td>
|
| 303 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">81.71</td>
|
| 304 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">85.37</td>
|
| 305 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">88.41</td>
|
| 306 |
+
</tr>
|
| 307 |
+
<tr>
|
| 308 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">HumanEval+</td>
|
| 309 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">pass@1</td>
|
| 310 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">76.83</td>
|
| 311 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">79.88</td>
|
| 312 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">85.37</td>
|
| 313 |
+
</tr>
|
| 314 |
+
<tr>
|
| 315 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MBPP</td>
|
| 316 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">pass@1</td>
|
| 317 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">71.16</td>
|
| 318 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">87.30</td>
|
| 319 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">85.45</td>
|
| 320 |
+
</tr>
|
| 321 |
+
<tr>
|
| 322 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MBPP+</td>
|
| 323 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">pass@1</td>
|
| 324 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">62.17</td>
|
| 325 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">73.81</td>
|
| 326 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">73.54</td>
|
| 327 |
+
</tr>
|
| 328 |
+
<tr>
|
| 329 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">CRUXEval-O</td>
|
| 330 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">pass@1</td>
|
| 331 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">40.75</td>
|
| 332 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">47.63</td>
|
| 333 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">55.75</td>
|
| 334 |
+
</tr>
|
| 335 |
+
<tr>
|
| 336 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">BigCodeBench</td>
|
| 337 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">pass@1</td>
|
| 338 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">32.19</td>
|
| 339 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">35.00</td>
|
| 340 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">38.77</td>
|
| 341 |
+
</tr>
|
| 342 |
+
<tr>
|
| 343 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MULTIPLE</td>
|
| 344 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">pass@1</td>
|
| 345 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">52.54</td>
|
| 346 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">60.26</td>
|
| 347 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">62.31</td>
|
| 348 |
+
</tr>
|
| 349 |
+
<tr>
|
| 350 |
+
<tr>
|
| 351 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">Eval+ Avg</td>
|
| 352 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">pass@1</td>
|
| 353 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">67.05</td>
|
| 354 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">80.21</td>
|
| 355 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">82.66</td>
|
| 356 |
+
</tr>
|
| 357 |
+
<tr>
|
| 358 |
+
<td colspan="5" style="text-align:center; background-color: #FFFFFF; color: #2D2D2D; font-style:italic;">
|
| 359 |
+
Tool Calling Tasks
|
| 360 |
+
</td>
|
| 361 |
+
</tr>
|
| 362 |
+
<tr>
|
| 363 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">BFCL v3</td>
|
| 364 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 365 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">60.80</td>
|
| 366 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">68.27</td>
|
| 367 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">73.68</td>
|
| 368 |
+
</tr>
|
| 369 |
+
<tr>
|
| 370 |
+
<td colspan="5" style="text-align:center; background-color: #FFFFFF; color: #2D2D2D; font-style:italic;">
|
| 371 |
+
Multilingual Tasks
|
| 372 |
+
</td>
|
| 373 |
+
</tr>
|
| 374 |
+
<tr>
|
| 375 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MMMLU</td>
|
| 376 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">5-shot</td>
|
| 377 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">57.61</td>
|
| 378 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">64.84</td>
|
| 379 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">73.71</td>
|
| 380 |
+
</tr>
|
| 381 |
+
<tr>
|
| 382 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">INCLUDE</td>
|
| 383 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">5-shot</td>
|
| 384 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">52.05</td>
|
| 385 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">58.89</td>
|
| 386 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">67.26</td>
|
| 387 |
+
</tr>
|
| 388 |
+
<tr>
|
| 389 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MGSM</td>
|
| 390 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">8-shot</td>
|
| 391 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">70.00</td>
|
| 392 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">82.32</td>
|
| 393 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">71.12</td>
|
| 394 |
+
</tr>
|
| 395 |
+
<tr>
|
| 396 |
+
<td colspan="6" style="text-align:center; background-color: #FFFFFF; color: #2D2D2D; font-style:italic;">
|
| 397 |
+
Safety
|
| 398 |
+
</td>
|
| 399 |
+
</tr>
|
| 400 |
+
<tr>
|
| 401 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">SALAD-Bench</td>
|
| 402 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 403 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">93.95</td>
|
| 404 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">95.80</td>
|
| 405 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">96.41</td>
|
| 406 |
+
</tr>
|
| 407 |
+
<tr>
|
| 408 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">AttaQ</td>
|
| 409 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 410 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">81.88</td>
|
| 411 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">81.19</td>
|
| 412 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">85.76</td>
|
| 413 |
+
</tr>
|
| 414 |
+
<tr>
|
| 415 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">Tulu3 Safety Eval Avg</td>
|
| 416 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;"></td>
|
| 417 |
+
<td style="text-align:right; background-color: #DAE8FF; color: #2D2D2D;">66.84</td>
|
| 418 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">75.57</td>
|
| 419 |
+
<td style="text-align:right; background-color: #FFFFFF; color: #2D2D2D;">78.19</td>
|
| 420 |
+
</tr>
|
| 421 |
+
</tbody></table>
|
| 422 |
+
|
| 423 |
+
|
| 424 |
+
<table>
|
| 425 |
+
<caption><b>Multilingual Benchmarks and the included languages:</b></caption>
|
| 426 |
+
<thead>
|
| 427 |
+
<tr>
|
| 428 |
+
<th style="text-align:left; background-color: #001d6c; color: white;">Benchmarks</th>
|
| 429 |
+
<th style="text-align:left; background-color: #001d6c; color: white;"># Langs</th>
|
| 430 |
+
<th style="text-align:center; background-color: #001d6c; color: white;">Languages</th>
|
| 431 |
+
</tr>
|
| 432 |
+
</thead>
|
| 433 |
+
<tbody>
|
| 434 |
+
<tr>
|
| 435 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MMMLU</td>
|
| 436 |
+
<td style="text-align:center; background-color: #FFFFFF; color: #2D2D2D;">11</td>
|
| 437 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">ar, de, en, es, fr, ja, ko, pt, zh, bn, hi</td>
|
| 438 |
+
</tr>
|
| 439 |
+
<tr>
|
| 440 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">INCLUDE</td>
|
| 441 |
+
<td style="text-align:center; background-color: #FFFFFF; color: #2D2D2D;">14</td>
|
| 442 |
+
<!-- <td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">hindi, bengali, tamil, telugu, arabic, german, spanish, french, italian, japanese, korean, dutch, portuguese, chinese</td> -->
|
| 443 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">hi, bn, ta, te, ar, de, es, fr, it, ja, ko, nl, pt, zh</td>
|
| 444 |
+
|
| 445 |
+
</tr>
|
| 446 |
+
<tr>
|
| 447 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">MGSM</td>
|
| 448 |
+
<td style="text-align:center; background-color: #FFFFFF; color: #2D2D2D;">5</td>
|
| 449 |
+
<td style="text-align:left; background-color: #FFFFFF; color: #2D2D2D;">en, es, fr, ja, zh</td>
|
| 450 |
+
</tr>
|
| 451 |
+
</tbody>
|
| 452 |
+
</table>
|
| 453 |
+
|
| 454 |
+
**Model Architecture:**
|
| 455 |
+
|
| 456 |
+
Granite-4.1-3B baseline is built on a decoder-only dense transformer architecture. Core components of this architecture are: GQA, RoPE, MLP with SwiGLU, RMSNorm, and shared input/output embeddings.
|
| 457 |
+
|
| 458 |
+
<table>
|
| 459 |
+
<thead>
|
| 460 |
+
<tr>
|
| 461 |
+
<th style="text-align:left; background-color: #001d6c; color: white;">Model</th>
|
| 462 |
+
<th style="text-align:center; background-color: #001d6c; color: white;">3B Dense</th>
|
| 463 |
+
<th style="text-align:center; background-color: #001d6c; color: white;">8B Dense</th>
|
| 464 |
+
<th style="text-align:center; background-color: #001d6c; color: white;">30B Dense</th>
|
| 465 |
+
</tr></thead>
|
| 466 |
+
<tbody>
|
| 467 |
+
<tr>
|
| 468 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Embedding size</td>
|
| 469 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">2560</td>
|
| 470 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">4096</td>
|
| 471 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">4096</td>
|
| 472 |
+
</tr>
|
| 473 |
+
<tr>
|
| 474 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Number of layers</td>
|
| 475 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">40</td>
|
| 476 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">40</td>
|
| 477 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">64</td>
|
| 478 |
+
</tr>
|
| 479 |
+
<tr>
|
| 480 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Attention head size</td>
|
| 481 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">64</td>
|
| 482 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">128</td>
|
| 483 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">128</td>
|
| 484 |
+
</tr>
|
| 485 |
+
<tr>
|
| 486 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Number of attention heads</td>
|
| 487 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">40</td>
|
| 488 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">32</td>
|
| 489 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">32</td>
|
| 490 |
+
</tr>
|
| 491 |
+
<tr>
|
| 492 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Number of KV heads</td>
|
| 493 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">8</td>
|
| 494 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">8</td>
|
| 495 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">8</td>
|
| 496 |
+
</tr>
|
| 497 |
+
<!--<tr>
|
| 498 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Mamba2 state size</td>
|
| 499 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">-</td>
|
| 500 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;"></td>
|
| 501 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 502 |
+
</tr>
|
| 503 |
+
<tr>
|
| 504 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Number of Mamba2 heads</td>
|
| 505 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 506 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;"></td>
|
| 507 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 508 |
+
</tr>-->
|
| 509 |
+
|
| 510 |
+
<tr>
|
| 511 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">MLP / Shared expert hidden size</td>
|
| 512 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">8192</td>
|
| 513 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">12800</td>
|
| 514 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">32768</td>
|
| 515 |
+
</tr>
|
| 516 |
+
<!--<tr>
|
| 517 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Num. Experts</td>
|
| 518 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 519 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;"></td>
|
| 520 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 521 |
+
</tr>
|
| 522 |
+
<tr>
|
| 523 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Num. active Experts</td>
|
| 524 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 525 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;"></td>
|
| 526 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 527 |
+
</tr>
|
| 528 |
+
<tr>
|
| 529 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Expert hidden size</td>
|
| 530 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 531 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;"></td>
|
| 532 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 533 |
+
</tr>-->
|
| 534 |
+
|
| 535 |
+
<tr>
|
| 536 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">MLP activation</td>
|
| 537 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">SwiGLU</td>
|
| 538 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">SwiGLU</td>
|
| 539 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">SwiGLU</td>
|
| 540 |
+
</tr>
|
| 541 |
+
|
| 542 |
+
<tr>
|
| 543 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Sequence length</td>
|
| 544 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">131072</td>
|
| 545 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">131072</td>
|
| 546 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">131072</td>
|
| 547 |
+
</tr>
|
| 548 |
+
<tr>
|
| 549 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;">Position embedding</td>
|
| 550 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">RoPE</td>
|
| 551 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">RoPE</td>
|
| 552 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">RoPE</td>
|
| 553 |
+
</tr>
|
| 554 |
+
<tr>
|
| 555 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;"># Parameters</td>
|
| 556 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;">3B</td>
|
| 557 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">8B</td>
|
| 558 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;">30B</td>
|
| 559 |
+
</tr>
|
| 560 |
+
<!-- <tr>
|
| 561 |
+
<td style="text-align:left; background-color: #FFFFFF; color: black;"># Active parameters</td>
|
| 562 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 563 |
+
<td style="text-align:center; background-color: #DAE8FF; color: black;"></td>
|
| 564 |
+
<td style="text-align:center; background-color: #FFFFFF; color: black;"></td>
|
| 565 |
+
</tr>-->
|
| 566 |
+
</tbody></table>
|
| 567 |
+
|
| 568 |
+
|
| 569 |
+
|
| 570 |
+
**Training Data:**
|
| 571 |
+
Overall, our SFT data is largely comprised of three key sources: (1) publicly available datasets with permissive license, (2) internal synthetic data targeting specific capabilities, and (3) a select set of human-curated data.
|
| 572 |
+
|
| 573 |
+
**Supervised Fine-Tuning and Reinforcement Learning:**
|
| 574 |
+
Instruct model has been fine tuned with significantly improved SFT-pipeline and Reinforcement learning pipelines with high quality mix of various datasets as mentioned above. With rigorous SFT-RL cycles we have improved Granite-4.1 model's tool calling, instruction following and chat capabilities. For further details please check our [Granite-4.1 Blog]((https://huggingface.co/blog/ibm-granite/granite-4-1)).
|
| 575 |
+
|
| 576 |
+
**Infrastructure:**
|
| 577 |
+
We trained the Granite 4.1 Language Models utilizing an NVIDIA GB200 NVL72 cluster hosted in CoreWeave. Intra-rack communication occurs via the 72-GPU NVLink domain, and a non-blocking, full Fat-Tree NDR 400 Gb/s InfiniBand network provides inter-rack communication. This cluster provides a scalable and efficient infrastructure for training our models over thousands of GPUs.
|
| 578 |
+
|
| 579 |
+
**Ethical Considerations and Limitations:**
|
| 580 |
+
Granite 4.1 Instruction Models are primarily finetuned using instruction-response pairs mostly in English, but also multilingual data covering multiple languages. Although this model can handle multilingual dialog use cases, its performance might not be similar to English tasks. In such cases, introducing a small number of examples (few-shot) can help the model in generating more accurate outputs. While this model has been aligned by keeping safety in consideration, the model may in some cases produce inaccurate, biased, or unsafe responses to user prompts. We urge the community to use this model with proper safety testing and tuning tailored for their specific tasks. To enhance safety in enterprise deployments, we recommend using Granite 4.1 Language models alongside [Granite Guardian](https://huggingface.co/ibm-granite/granite-guardian-4.1-8b), a model designed to detect and flag risks in inputs and outputs across key dimensions outlined in the IBM AI Risk Atlas.
|
| 581 |
+
|
| 582 |
+
**Resources**
|
| 583 |
+
- ⭐️ Learn about the latest updates with Granite: https://www.ibm.com/granite
|
| 584 |
+
- 📄 Get started with tutorials, best practices, and prompt engineering advice: https://www.ibm.com/granite/docs/
|
| 585 |
+
- 💡 Learn about the latest Granite learning resources: https://ibm.biz/granite-learning-resources
|
| 586 |
+
|
| 587 |
+
<!-- ## Citation
|
| 588 |
+
```
|
| 589 |
+
@misc{granite-models,
|
| 590 |
+
author = {author 1, author2, ...},
|
| 591 |
+
title = {},
|
| 592 |
+
journal = {},
|
| 593 |
+
volume = {},
|
| 594 |
+
year = {2024},
|
| 595 |
+
url = {https://arxiv.org/abs/0000.00000},
|
| 596 |
+
}
|
| 597 |
+
``` -->
|
adapter_index.json
ADDED
|
@@ -0,0 +1,151 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_info": {
|
| 3 |
+
"num_adapters": 13,
|
| 4 |
+
"base_model": "granite-4.1-3b"
|
| 5 |
+
},
|
| 6 |
+
"adapters": [
|
| 7 |
+
{
|
| 8 |
+
"adapter_index": 1,
|
| 9 |
+
"adapter_name": "citations",
|
| 10 |
+
"technology": "lora",
|
| 11 |
+
"control_token": {
|
| 12 |
+
"token": "<|citations|>",
|
| 13 |
+
"id": 100352
|
| 14 |
+
},
|
| 15 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-rag-r1.0/snapshots/6e4a75e35f1cb272e8d15b4615fb0a123398d1cf/citations/granite-4.1-3b/lora",
|
| 16 |
+
"io_config": "io_configs/citations/io.yaml"
|
| 17 |
+
},
|
| 18 |
+
{
|
| 19 |
+
"adapter_index": 2,
|
| 20 |
+
"adapter_name": "query_rewrite",
|
| 21 |
+
"technology": "lora",
|
| 22 |
+
"control_token": {
|
| 23 |
+
"token": "<|query_rewrite|>",
|
| 24 |
+
"id": 100353
|
| 25 |
+
},
|
| 26 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-rag-r1.0/snapshots/6e4a75e35f1cb272e8d15b4615fb0a123398d1cf/query_rewrite/granite-4.1-3b/lora",
|
| 27 |
+
"io_config": "io_configs/query_rewrite/io.yaml"
|
| 28 |
+
},
|
| 29 |
+
{
|
| 30 |
+
"adapter_index": 3,
|
| 31 |
+
"adapter_name": "query_clarification",
|
| 32 |
+
"technology": "lora",
|
| 33 |
+
"control_token": {
|
| 34 |
+
"token": "<|query_clarification|>",
|
| 35 |
+
"id": 100354
|
| 36 |
+
},
|
| 37 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-rag-r1.0/snapshots/6e4a75e35f1cb272e8d15b4615fb0a123398d1cf/query_clarification/granite-4.1-3b/lora",
|
| 38 |
+
"io_config": "io_configs/query_clarification/io.yaml"
|
| 39 |
+
},
|
| 40 |
+
{
|
| 41 |
+
"adapter_index": 4,
|
| 42 |
+
"adapter_name": "context_relevance",
|
| 43 |
+
"technology": "lora",
|
| 44 |
+
"control_token": {
|
| 45 |
+
"token": "<|context_relevance|>",
|
| 46 |
+
"id": 100355
|
| 47 |
+
},
|
| 48 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-rag-r1.0/snapshots/6e4a75e35f1cb272e8d15b4615fb0a123398d1cf/context_relevance/granite-4.1-3b/lora",
|
| 49 |
+
"io_config": "io_configs/context_relevance/io.yaml"
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"adapter_index": 5,
|
| 53 |
+
"adapter_name": "hallucination_detection",
|
| 54 |
+
"technology": "lora",
|
| 55 |
+
"control_token": {
|
| 56 |
+
"token": "<|hallucination_detection|>",
|
| 57 |
+
"id": 100356
|
| 58 |
+
},
|
| 59 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-rag-r1.0/snapshots/6e4a75e35f1cb272e8d15b4615fb0a123398d1cf/hallucination_detection/granite-4.1-3b/lora",
|
| 60 |
+
"io_config": "io_configs/hallucination_detection/io.yaml"
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"adapter_index": 6,
|
| 64 |
+
"adapter_name": "answerability",
|
| 65 |
+
"technology": "lora",
|
| 66 |
+
"control_token": {
|
| 67 |
+
"token": "<|answerability|>",
|
| 68 |
+
"id": 100357
|
| 69 |
+
},
|
| 70 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-rag-r1.0/snapshots/6e4a75e35f1cb272e8d15b4615fb0a123398d1cf/answerability/granite-4.1-3b/lora",
|
| 71 |
+
"io_config": "io_configs/answerability/io.yaml"
|
| 72 |
+
},
|
| 73 |
+
{
|
| 74 |
+
"adapter_index": 7,
|
| 75 |
+
"adapter_name": "uncertainty",
|
| 76 |
+
"technology": "lora",
|
| 77 |
+
"control_token": {
|
| 78 |
+
"token": "<|uncertainty|>",
|
| 79 |
+
"id": 100358
|
| 80 |
+
},
|
| 81 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-core-r1.0/snapshots/8f78babf3f0d5baba230464838050a71fe59dee5/uncertainty/granite-4.1-3b/lora",
|
| 82 |
+
"io_config": "io_configs/uncertainty/io.yaml"
|
| 83 |
+
},
|
| 84 |
+
{
|
| 85 |
+
"adapter_index": 8,
|
| 86 |
+
"adapter_name": "requirement-check",
|
| 87 |
+
"technology": "lora",
|
| 88 |
+
"control_token": {
|
| 89 |
+
"token": "<|requirement-check|>",
|
| 90 |
+
"id": 100359
|
| 91 |
+
},
|
| 92 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-core-r1.0/snapshots/8f78babf3f0d5baba230464838050a71fe59dee5/requirement-check/granite-4.1-3b/lora",
|
| 93 |
+
"io_config": "io_configs/requirement-check/io.yaml"
|
| 94 |
+
},
|
| 95 |
+
{
|
| 96 |
+
"adapter_index": 9,
|
| 97 |
+
"adapter_name": "context-attribution",
|
| 98 |
+
"technology": "lora",
|
| 99 |
+
"control_token": {
|
| 100 |
+
"token": "<|context-attribution|>",
|
| 101 |
+
"id": 100360
|
| 102 |
+
},
|
| 103 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-core-r1.0/snapshots/8f78babf3f0d5baba230464838050a71fe59dee5/context-attribution/granite-4.1-3b/lora",
|
| 104 |
+
"io_config": "io_configs/context-attribution/io.yaml"
|
| 105 |
+
},
|
| 106 |
+
{
|
| 107 |
+
"adapter_index": 10,
|
| 108 |
+
"adapter_name": "factuality-detection",
|
| 109 |
+
"technology": "lora",
|
| 110 |
+
"control_token": {
|
| 111 |
+
"token": "<|factuality-detection|>",
|
| 112 |
+
"id": 100361
|
| 113 |
+
},
|
| 114 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-guardian-r1.0/snapshots/882ccf11cf1e4cdc3a66044f17872e55078dbc85/factuality-detection/granite-4.1-3b/lora",
|
| 115 |
+
"io_config": "io_configs/factuality-detection/io.yaml"
|
| 116 |
+
},
|
| 117 |
+
{
|
| 118 |
+
"adapter_index": 11,
|
| 119 |
+
"adapter_name": "policy-guardrails",
|
| 120 |
+
"technology": "lora",
|
| 121 |
+
"control_token": {
|
| 122 |
+
"token": "<|policy-guardrails|>",
|
| 123 |
+
"id": 100362
|
| 124 |
+
},
|
| 125 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-guardian-r1.0/snapshots/882ccf11cf1e4cdc3a66044f17872e55078dbc85/policy-guardrails/granite-4.1-3b/lora",
|
| 126 |
+
"io_config": "io_configs/policy-guardrails/io.yaml"
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"adapter_index": 12,
|
| 130 |
+
"adapter_name": "factuality-correction",
|
| 131 |
+
"technology": "lora",
|
| 132 |
+
"control_token": {
|
| 133 |
+
"token": "<|factuality-correction|>",
|
| 134 |
+
"id": 100363
|
| 135 |
+
},
|
| 136 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-guardian-r1.0/snapshots/882ccf11cf1e4cdc3a66044f17872e55078dbc85/factuality-correction/granite-4.1-3b/lora",
|
| 137 |
+
"io_config": "io_configs/factuality-correction/io.yaml"
|
| 138 |
+
},
|
| 139 |
+
{
|
| 140 |
+
"adapter_index": 13,
|
| 141 |
+
"adapter_name": "guardian-core",
|
| 142 |
+
"technology": "lora",
|
| 143 |
+
"control_token": {
|
| 144 |
+
"token": "<|guardian-core|>",
|
| 145 |
+
"id": 100364
|
| 146 |
+
},
|
| 147 |
+
"original_path": "/proj/dmfexp/lastrasl/hfhub/hub/models--ibm-granite--granitelib-guardian-r1.0/snapshots/882ccf11cf1e4cdc3a66044f17872e55078dbc85/guardian-core/granite-4.1-3b/lora",
|
| 148 |
+
"io_config": "io_configs/guardian-core/io.yaml"
|
| 149 |
+
}
|
| 150 |
+
]
|
| 151 |
+
}
|
chat_template.jinja
ADDED
|
@@ -0,0 +1,185 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{%- set adapter_map = {
|
| 2 |
+
'citations': {'token': '<|citations|>', 'type': 'lora'},
|
| 3 |
+
'query_rewrite': {'token': '<|query_rewrite|>', 'type': 'lora'},
|
| 4 |
+
'query_clarification': {'token': '<|query_clarification|>', 'type': 'lora'},
|
| 5 |
+
'context_relevance': {'token': '<|context_relevance|>', 'type': 'lora'},
|
| 6 |
+
'hallucination_detection': {'token': '<|hallucination_detection|>', 'type': 'lora'},
|
| 7 |
+
'answerability': {'token': '<|answerability|>', 'type': 'lora'},
|
| 8 |
+
'uncertainty': {'token': '<|uncertainty|>', 'type': 'lora'},
|
| 9 |
+
'requirement-check': {'token': '<|requirement-check|>', 'type': 'lora'},
|
| 10 |
+
'context-attribution': {'token': '<|context-attribution|>', 'type': 'lora'},
|
| 11 |
+
'factuality-detection': {'token': '<|factuality-detection|>', 'type': 'lora'},
|
| 12 |
+
'policy-guardrails': {'token': '<|policy-guardrails|>', 'type': 'lora'},
|
| 13 |
+
'factuality-correction': {'token': '<|factuality-correction|>', 'type': 'lora'},
|
| 14 |
+
'guardian-core': {'token': '<|guardian-core|>', 'type': 'lora'}
|
| 15 |
+
} %}
|
| 16 |
+
{#- Look up adapter token, type, and invocation text from adapter_name -#}
|
| 17 |
+
{%- set adapter_token = '' %}
|
| 18 |
+
{%- set adapter_type = '' %}
|
| 19 |
+
{%- set adapter_invocation_text = '' %}
|
| 20 |
+
{%- if adapter_name is defined and adapter_name in adapter_map %}
|
| 21 |
+
{%- set adapter_token = adapter_map[adapter_name]['token'] %}
|
| 22 |
+
{%- set adapter_type = adapter_map[adapter_name]['type'] %}
|
| 23 |
+
{%- if adapter_map[adapter_name]['type'] == 'alora' %}
|
| 24 |
+
{%- set adapter_invocation_text = adapter_map[adapter_name]['invocation_text'] %}
|
| 25 |
+
{%- endif %}
|
| 26 |
+
{%- endif %}
|
| 27 |
+
|
| 28 |
+
{%- set tools_system_message_prefix = 'You are a helpful assistant with access to the following tools. You may call one or more tools to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>' %}
|
| 29 |
+
{%- set tools_system_message_suffix = '\n</tools>\n\nFor each tool call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call>. If a tool does not exist in the provided list of tools, notify the user that you do not have the ability to fulfill the request.' %}
|
| 30 |
+
{%- set documents_system_message_prefix = 'You are a helpful assistant with access to the following documents. You may use one or more documents to assist with the user query.\n\nYou are given a list of documents within <documents></documents> XML tags:\n<documents>' %}
|
| 31 |
+
{%- set documents_system_message_suffix = '\n</documents>\n\nWrite the response to the user\'s input by strictly aligning with the facts in the provided documents. If the information needed to answer the question is not available in the documents, inform the user that the question cannot be answered based on the available data.' %}
|
| 32 |
+
{%- if available_tools is defined and available_tools %}
|
| 33 |
+
{%- set tools = available_tools %}
|
| 34 |
+
{%- endif %}
|
| 35 |
+
{%- set ns = namespace(tools_system_message=tools_system_message_prefix,
|
| 36 |
+
documents_system_message=documents_system_message_prefix,
|
| 37 |
+
system_message=''
|
| 38 |
+
,
|
| 39 |
+
adapter_token=adapter_token,
|
| 40 |
+
adapter_type=adapter_type,
|
| 41 |
+
adapter_invocation_text=adapter_invocation_text,
|
| 42 |
+
alora_target_idx=-1
|
| 43 |
+
) %}
|
| 44 |
+
{%- if tools %}
|
| 45 |
+
{%- for tool in tools %}
|
| 46 |
+
{%- set ns.tools_system_message = ns.tools_system_message + '\n' + (tool | tojson) %}
|
| 47 |
+
{%- endfor %}
|
| 48 |
+
{%- set ns.tools_system_message = ns.tools_system_message + tools_system_message_suffix %}
|
| 49 |
+
{%- else %}
|
| 50 |
+
{%- set ns.tools_system_message = '' %}
|
| 51 |
+
{%- endif %}
|
| 52 |
+
{%- if documents %}
|
| 53 |
+
{%- for document in documents %}
|
| 54 |
+
{%- set ns.documents_system_message = ns.documents_system_message + '\n' + (document | tojson) %}
|
| 55 |
+
{%- endfor %}
|
| 56 |
+
{%- set ns.documents_system_message = ns.documents_system_message + documents_system_message_suffix %}
|
| 57 |
+
{%- else %}
|
| 58 |
+
{%- set ns.documents_system_message = '' %}
|
| 59 |
+
{%- endif %}
|
| 60 |
+
{#- For lora adapters: insert activation token at the very beginning -#}
|
| 61 |
+
{%- if ns.adapter_token and ns.adapter_type == 'lora' %}
|
| 62 |
+
{{- ns.adapter_token }}
|
| 63 |
+
{%- endif %}
|
| 64 |
+
|
| 65 |
+
{%- if messages[0].role == 'system' %}
|
| 66 |
+
{%- if messages[0].content is string %}
|
| 67 |
+
{%- set ns.system_message = messages[0].content %}
|
| 68 |
+
{%- elif messages[0].content is iterable %}
|
| 69 |
+
{%- for entry in messages[0].content %}
|
| 70 |
+
{%- if entry.type== 'text' %}
|
| 71 |
+
{%- if ns.system_message != '' %}
|
| 72 |
+
{%- set ns.system_message = ns.system_message + '\n' %}
|
| 73 |
+
{%- endif %}
|
| 74 |
+
{%- set ns.system_message = ns.system_message + entry.text %}
|
| 75 |
+
{%- endif %}
|
| 76 |
+
{%- endfor %}
|
| 77 |
+
{%- endif %}
|
| 78 |
+
{%- if tools and documents %}
|
| 79 |
+
{%- set ns.system_message = ns.system_message + '\n\n' + ns.tools_system_message + '\n\n' + ns.documents_system_message %}
|
| 80 |
+
{%- elif tools %}
|
| 81 |
+
{%- set ns.system_message = ns.system_message + '\n\n' + ns.tools_system_message %}
|
| 82 |
+
{%- elif documents %}
|
| 83 |
+
{%- set ns.system_message = ns.system_message + '\n\n' + ns.documents_system_message %}
|
| 84 |
+
{%- endif %}
|
| 85 |
+
{%- else %}
|
| 86 |
+
{%- if tools and documents %}
|
| 87 |
+
{%- set ns.system_message = ns.tools_system_message + '\n\n' + ns.documents_system_message %}
|
| 88 |
+
{%- elif tools %}
|
| 89 |
+
{%- set ns.system_message = ns.tools_system_message %}
|
| 90 |
+
{%- elif documents %}
|
| 91 |
+
{%- set ns.system_message = ns.documents_system_message %}
|
| 92 |
+
{%- endif %}
|
| 93 |
+
{%- endif %}
|
| 94 |
+
{%- if ns.system_message %}
|
| 95 |
+
{{- '<|start_of_role|>system<|end_of_role|>' + ns.system_message + '<|end_of_text|>\n' }}
|
| 96 |
+
{%- endif %}
|
| 97 |
+
{#- ALoRA Pass 1: find the last user message containing the invocation text.
|
| 98 |
+
ns.alora_target_idx stays -1 when the invocation sequence is the assistant role
|
| 99 |
+
token sequence (not present in any user message); the fallback insertion below
|
| 100 |
+
handles that case. -#}
|
| 101 |
+
{%- if ns.adapter_type == 'alora' and ns.adapter_invocation_text %}
|
| 102 |
+
{%- for _msg in messages %}
|
| 103 |
+
{%- if _msg.role == 'user' %}
|
| 104 |
+
{%- if _msg.content is string and ns.adapter_invocation_text in _msg.content %}
|
| 105 |
+
{%- set ns.alora_target_idx = loop.index0 %}
|
| 106 |
+
{%- elif _msg.content is not string and _msg.content is iterable %}
|
| 107 |
+
{%- set _msg_idx = loop.index0 %}
|
| 108 |
+
{%- for _entry in _msg.content %}
|
| 109 |
+
{%- if _entry.type == 'text' and ns.adapter_invocation_text in _entry.text %}
|
| 110 |
+
{%- set ns.alora_target_idx = _msg_idx %}
|
| 111 |
+
{%- endif %}
|
| 112 |
+
{%- endfor %}
|
| 113 |
+
{%- endif %}
|
| 114 |
+
{%- endif %}
|
| 115 |
+
{%- endfor %}
|
| 116 |
+
{%- endif %}
|
| 117 |
+
{%- for message in messages %}
|
| 118 |
+
{%- set content = namespace(val='') %}
|
| 119 |
+
{%- if message.content is string %}
|
| 120 |
+
{%- set content.val = message.content %}
|
| 121 |
+
{%- else %}
|
| 122 |
+
{%- if message.content is iterable %}
|
| 123 |
+
{%- for entry in message.content %}
|
| 124 |
+
{%- if entry.type== 'text' %}
|
| 125 |
+
{%- if content.val != '' %}
|
| 126 |
+
{%- set content.val = content.val + '\n' %}
|
| 127 |
+
{%- endif %}
|
| 128 |
+
{%- set content.val = content.val + entry.text %}
|
| 129 |
+
{%- endif %}
|
| 130 |
+
{%- endfor %}
|
| 131 |
+
{%- endif %}
|
| 132 |
+
{%- endif %}
|
| 133 |
+
{#- ALoRA Pass 2: inject activation token before invocation text in the target message -#}
|
| 134 |
+
{%- if loop.index0 == ns.alora_target_idx %}
|
| 135 |
+
{%- set _parts = content.val.rsplit(ns.adapter_invocation_text, 1) %}
|
| 136 |
+
{%- if _parts | length > 1 %}
|
| 137 |
+
{%- set content.val = _parts[0] + ns.adapter_token + ns.adapter_invocation_text + _parts[1] %}
|
| 138 |
+
{%- endif %}
|
| 139 |
+
{%- endif %}
|
| 140 |
+
{%- if (message.role == 'user') or (message.role == 'system' and not loop.first) %}
|
| 141 |
+
{{- '<|start_of_role|>' + message.role + '<|end_of_role|>' + content.val + '<|end_of_text|>\n' }}
|
| 142 |
+
{%- elif message.role == 'assistant' %}
|
| 143 |
+
{{- '<|start_of_role|>' + message.role + '<|end_of_role|>' + content.val }}
|
| 144 |
+
{%- if message.tool_calls %}
|
| 145 |
+
{%- for tool_call in message.tool_calls %}
|
| 146 |
+
{%- if (loop.first and content.val) or (not loop.first) %}
|
| 147 |
+
{{- '\n' }}
|
| 148 |
+
{%- endif %}
|
| 149 |
+
{%- if tool_call.function %}
|
| 150 |
+
{%- set tool_call = tool_call.function %}
|
| 151 |
+
{%- endif %}
|
| 152 |
+
{{- '<tool_call>\n{"name": "' }}
|
| 153 |
+
{{- tool_call.name }}
|
| 154 |
+
{{- '", "arguments": ' }}
|
| 155 |
+
{%- if tool_call.arguments is string %}
|
| 156 |
+
{{- tool_call.arguments }}
|
| 157 |
+
{%- else %}
|
| 158 |
+
{{- tool_call.arguments | tojson }}
|
| 159 |
+
{%- endif %}
|
| 160 |
+
{{- '}\n</tool_call>' }}
|
| 161 |
+
{%- endfor %}
|
| 162 |
+
{%- endif %}
|
| 163 |
+
{{- '<|end_of_text|>\n' }}
|
| 164 |
+
{%- elif message.role == 'tool' %}
|
| 165 |
+
{%- if loop.first or (messages[loop.index0 - 1].role != 'tool') %}
|
| 166 |
+
{{- '<|start_of_role|>user<|end_of_role|>' }}
|
| 167 |
+
{%- endif %}
|
| 168 |
+
{{- '\n<tool_response>\n' }}
|
| 169 |
+
{{- content.val }}
|
| 170 |
+
{{- '\n</tool_response>' }}
|
| 171 |
+
{%- if loop.last or (messages[loop.index0 + 1].role != 'tool') %}
|
| 172 |
+
{{- '<|end_of_text|>\n' }}
|
| 173 |
+
{%- endif %}
|
| 174 |
+
{%- endif %}
|
| 175 |
+
{%- endfor %}
|
| 176 |
+
{#- ALoRA fallback: insert activation token right before generation prompt.
|
| 177 |
+
Only fires when Pass 1 found no user message with the invocation text
|
| 178 |
+
(alora_target_idx == -1), meaning the adapter activates at the assistant
|
| 179 |
+
role token boundary rather than inside a user message. -#}
|
| 180 |
+
{%- if ns.adapter_token and ns.adapter_type == 'alora' and ns.alora_target_idx == -1 %}
|
| 181 |
+
{{- ns.adapter_token }}
|
| 182 |
+
{%- endif %}
|
| 183 |
+
{%- if add_generation_prompt %}
|
| 184 |
+
{{- '<|start_of_role|>assistant<|end_of_role|>' }}
|
| 185 |
+
{%- endif %}
|
compose_report.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
config.json
ADDED
|
@@ -0,0 +1,233 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"adapter_names": [
|
| 3 |
+
"citations",
|
| 4 |
+
"query_rewrite",
|
| 5 |
+
"query_clarification",
|
| 6 |
+
"context_relevance",
|
| 7 |
+
"hallucination_detection",
|
| 8 |
+
"answerability",
|
| 9 |
+
"uncertainty",
|
| 10 |
+
"requirement-check",
|
| 11 |
+
"context-attribution",
|
| 12 |
+
"factuality-detection",
|
| 13 |
+
"policy-guardrails",
|
| 14 |
+
"factuality-correction",
|
| 15 |
+
"guardian-core"
|
| 16 |
+
],
|
| 17 |
+
"adapter_ranks": [
|
| 18 |
+
16,
|
| 19 |
+
32,
|
| 20 |
+
32,
|
| 21 |
+
16,
|
| 22 |
+
16,
|
| 23 |
+
16,
|
| 24 |
+
32,
|
| 25 |
+
64,
|
| 26 |
+
16,
|
| 27 |
+
32,
|
| 28 |
+
16,
|
| 29 |
+
32,
|
| 30 |
+
32
|
| 31 |
+
],
|
| 32 |
+
"adapter_third_party": [
|
| 33 |
+
"citations",
|
| 34 |
+
"query_rewrite",
|
| 35 |
+
"query_clarification",
|
| 36 |
+
"context_relevance",
|
| 37 |
+
"hallucination_detection",
|
| 38 |
+
"answerability",
|
| 39 |
+
"uncertainty",
|
| 40 |
+
"requirement-check",
|
| 41 |
+
"context-attribution",
|
| 42 |
+
"factuality-detection",
|
| 43 |
+
"policy-guardrails",
|
| 44 |
+
"factuality-correction",
|
| 45 |
+
"guardian-core"
|
| 46 |
+
],
|
| 47 |
+
"adapter_token_ids": [
|
| 48 |
+
100352,
|
| 49 |
+
100353,
|
| 50 |
+
100354,
|
| 51 |
+
100355,
|
| 52 |
+
100356,
|
| 53 |
+
100357,
|
| 54 |
+
100358,
|
| 55 |
+
100359,
|
| 56 |
+
100360,
|
| 57 |
+
100361,
|
| 58 |
+
100362,
|
| 59 |
+
100363,
|
| 60 |
+
100364
|
| 61 |
+
],
|
| 62 |
+
"architectures": [
|
| 63 |
+
"GraniteSwitchForCausalLM"
|
| 64 |
+
],
|
| 65 |
+
"attention_bias": false,
|
| 66 |
+
"attention_dropout": 0.0,
|
| 67 |
+
"attention_multiplier": 0.015625,
|
| 68 |
+
"bos_token_id": 100257,
|
| 69 |
+
"control_dims": 32,
|
| 70 |
+
"control_token_gain": 15.0,
|
| 71 |
+
"dtype": "bfloat16",
|
| 72 |
+
"embedding_multiplier": 12.0,
|
| 73 |
+
"eos_token_id": 100257,
|
| 74 |
+
"fused_add_norm": false,
|
| 75 |
+
"hidden_act": "silu",
|
| 76 |
+
"hidden_size": 2560,
|
| 77 |
+
"hiding_groups": {
|
| 78 |
+
"all_controls": [
|
| 79 |
+
"citations",
|
| 80 |
+
"query_rewrite",
|
| 81 |
+
"query_clarification",
|
| 82 |
+
"context_relevance",
|
| 83 |
+
"hallucination_detection",
|
| 84 |
+
"answerability",
|
| 85 |
+
"uncertainty",
|
| 86 |
+
"requirement-check",
|
| 87 |
+
"context-attribution",
|
| 88 |
+
"factuality-detection",
|
| 89 |
+
"policy-guardrails",
|
| 90 |
+
"factuality-correction",
|
| 91 |
+
"guardian-core"
|
| 92 |
+
]
|
| 93 |
+
},
|
| 94 |
+
"hiding_policy": {
|
| 95 |
+
"answerability": [
|
| 96 |
+
"all_controls"
|
| 97 |
+
],
|
| 98 |
+
"base": [
|
| 99 |
+
"all_controls"
|
| 100 |
+
],
|
| 101 |
+
"citations": [
|
| 102 |
+
"all_controls"
|
| 103 |
+
],
|
| 104 |
+
"context-attribution": [
|
| 105 |
+
"all_controls"
|
| 106 |
+
],
|
| 107 |
+
"context_relevance": [
|
| 108 |
+
"all_controls"
|
| 109 |
+
],
|
| 110 |
+
"factuality-correction": [
|
| 111 |
+
"all_controls"
|
| 112 |
+
],
|
| 113 |
+
"factuality-detection": [
|
| 114 |
+
"all_controls"
|
| 115 |
+
],
|
| 116 |
+
"guardian-core": [
|
| 117 |
+
"all_controls"
|
| 118 |
+
],
|
| 119 |
+
"hallucination_detection": [
|
| 120 |
+
"all_controls"
|
| 121 |
+
],
|
| 122 |
+
"policy-guardrails": [
|
| 123 |
+
"all_controls"
|
| 124 |
+
],
|
| 125 |
+
"query_clarification": [
|
| 126 |
+
"all_controls"
|
| 127 |
+
],
|
| 128 |
+
"query_rewrite": [
|
| 129 |
+
"all_controls"
|
| 130 |
+
],
|
| 131 |
+
"requirement-check": [
|
| 132 |
+
"all_controls"
|
| 133 |
+
],
|
| 134 |
+
"uncertainty": [
|
| 135 |
+
"all_controls"
|
| 136 |
+
]
|
| 137 |
+
},
|
| 138 |
+
"initializer_range": 0.1,
|
| 139 |
+
"intermediate_size": 8192,
|
| 140 |
+
"layer_types": [
|
| 141 |
+
"attention",
|
| 142 |
+
"attention",
|
| 143 |
+
"attention",
|
| 144 |
+
"attention",
|
| 145 |
+
"attention",
|
| 146 |
+
"attention",
|
| 147 |
+
"attention",
|
| 148 |
+
"attention",
|
| 149 |
+
"attention",
|
| 150 |
+
"attention",
|
| 151 |
+
"attention",
|
| 152 |
+
"attention",
|
| 153 |
+
"attention",
|
| 154 |
+
"attention",
|
| 155 |
+
"attention",
|
| 156 |
+
"attention",
|
| 157 |
+
"attention",
|
| 158 |
+
"attention",
|
| 159 |
+
"attention",
|
| 160 |
+
"attention",
|
| 161 |
+
"attention",
|
| 162 |
+
"attention",
|
| 163 |
+
"attention",
|
| 164 |
+
"attention",
|
| 165 |
+
"attention",
|
| 166 |
+
"attention",
|
| 167 |
+
"attention",
|
| 168 |
+
"attention",
|
| 169 |
+
"attention",
|
| 170 |
+
"attention",
|
| 171 |
+
"attention",
|
| 172 |
+
"attention",
|
| 173 |
+
"attention",
|
| 174 |
+
"attention",
|
| 175 |
+
"attention",
|
| 176 |
+
"attention",
|
| 177 |
+
"attention",
|
| 178 |
+
"attention",
|
| 179 |
+
"attention",
|
| 180 |
+
"attention",
|
| 181 |
+
"attention"
|
| 182 |
+
],
|
| 183 |
+
"logits_scaling": 10.0,
|
| 184 |
+
"lora_target_modules": [
|
| 185 |
+
"o_proj",
|
| 186 |
+
"qkv_proj",
|
| 187 |
+
"shared_input_linear",
|
| 188 |
+
"shared_output_linear"
|
| 189 |
+
],
|
| 190 |
+
"mamba_chunk_size": 256,
|
| 191 |
+
"mamba_conv_bias": true,
|
| 192 |
+
"mamba_d_conv": 4,
|
| 193 |
+
"mamba_d_head": 40,
|
| 194 |
+
"mamba_d_state": 256,
|
| 195 |
+
"mamba_expand": 2,
|
| 196 |
+
"mamba_n_groups": 1,
|
| 197 |
+
"mamba_n_heads": 128,
|
| 198 |
+
"mamba_proj_bias": false,
|
| 199 |
+
"max_lora_rank": 64,
|
| 200 |
+
"max_position_embeddings": 131072,
|
| 201 |
+
"model_type": "granite_switch",
|
| 202 |
+
"num_adapters": 13,
|
| 203 |
+
"num_attention_heads": 40,
|
| 204 |
+
"num_experts_per_tok": 2,
|
| 205 |
+
"num_hidden_layers": 41,
|
| 206 |
+
"num_key_value_heads": 8,
|
| 207 |
+
"num_local_experts": 0,
|
| 208 |
+
"output_router_logits": false,
|
| 209 |
+
"pad_token_id": 100256,
|
| 210 |
+
"position_embedding_type": "rope",
|
| 211 |
+
"projection_head_dim": 64,
|
| 212 |
+
"residual_multiplier": 0.22,
|
| 213 |
+
"rms_norm_eps": 1e-05,
|
| 214 |
+
"rope_parameters": {
|
| 215 |
+
"rope_theta": 10000000,
|
| 216 |
+
"rope_type": "default"
|
| 217 |
+
},
|
| 218 |
+
"router_aux_loss_coef": 0.001,
|
| 219 |
+
"shared_intermediate_size": 8192,
|
| 220 |
+
"switch_head_dim": 32,
|
| 221 |
+
"tie_word_embeddings": true,
|
| 222 |
+
"time_step_limit": [
|
| 223 |
+
0.0,
|
| 224 |
+
{
|
| 225 |
+
"__float__": "Infinity"
|
| 226 |
+
}
|
| 227 |
+
],
|
| 228 |
+
"time_step_max": 0.1,
|
| 229 |
+
"time_step_min": 0.001,
|
| 230 |
+
"transformers_version": "5.6.2",
|
| 231 |
+
"use_cache": true,
|
| 232 |
+
"vocab_size": 100365
|
| 233 |
+
}
|
generation_config.json
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 100257,
|
| 4 |
+
"eos_token_id": 100257,
|
| 5 |
+
"output_attentions": false,
|
| 6 |
+
"output_hidden_states": false,
|
| 7 |
+
"pad_token_id": 100256,
|
| 8 |
+
"transformers_version": "5.6.2",
|
| 9 |
+
"use_cache": true
|
| 10 |
+
}
|
io_configs/answerability/io.yaml
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: answerability
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request.
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"type": "string",
|
| 8 |
+
"enum": ["answerable", "unanswerable"]
|
| 9 |
+
}
|
| 10 |
+
transformations:
|
| 11 |
+
# Convert scalar value to a record for consistency with other intrinsics
|
| 12 |
+
- type: nest
|
| 13 |
+
input_path: []
|
| 14 |
+
field_name: "answerability"
|
| 15 |
+
instruction: ~
|
| 16 |
+
parameters:
|
| 17 |
+
temperature: 0.0
|
| 18 |
+
# "unanswerable" can be 6 tokens at high temperatures
|
| 19 |
+
max_completion_tokens: 6
|
| 20 |
+
# No sentence boundary detection
|
| 21 |
+
sentence_boundaries: ~
|
io_configs/citations/io.yaml
ADDED
|
@@ -0,0 +1,98 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: citations
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"$defs": {
|
| 8 |
+
"_MODEL_OUTPUT_ENTRY": {
|
| 9 |
+
"properties": {
|
| 10 |
+
"r": {
|
| 11 |
+
"minimum": 0,
|
| 12 |
+
"title": "R",
|
| 13 |
+
"type": "integer"
|
| 14 |
+
},
|
| 15 |
+
"c": {
|
| 16 |
+
"items": {
|
| 17 |
+
"minimum": 0,
|
| 18 |
+
"type": "integer"
|
| 19 |
+
},
|
| 20 |
+
"title": "C",
|
| 21 |
+
"type": "array"
|
| 22 |
+
}
|
| 23 |
+
},
|
| 24 |
+
"required": [
|
| 25 |
+
"r",
|
| 26 |
+
"c"
|
| 27 |
+
],
|
| 28 |
+
"title": "_MODEL_OUTPUT_ENTRY",
|
| 29 |
+
"type": "object"
|
| 30 |
+
}
|
| 31 |
+
},
|
| 32 |
+
"items": {
|
| 33 |
+
"$ref": "#/$defs/_MODEL_OUTPUT_ENTRY"
|
| 34 |
+
},
|
| 35 |
+
"title": "_MODEL_OUTPUT",
|
| 36 |
+
"type": "array"
|
| 37 |
+
}
|
| 38 |
+
transformations:
|
| 39 |
+
# Explode the list of document sentences in each citation
|
| 40 |
+
- type: explode
|
| 41 |
+
input_path: [] # Zero-length path means match root element
|
| 42 |
+
target_field: "c"
|
| 43 |
+
# Model may repeat itself; drop the resulting duplicates.
|
| 44 |
+
- type: drop_duplicates
|
| 45 |
+
input_path: [] # Zero-length path means match root element
|
| 46 |
+
target_fields: ["r", "c"]
|
| 47 |
+
# Replace sentence number with sentence location and contents.
|
| 48 |
+
# Do this first for sentences from the last turn, then for sentences from documents.
|
| 49 |
+
- type: decode_sentences
|
| 50 |
+
source: "last_message"
|
| 51 |
+
input_path: [~, "r"] # Null in path means wildcard
|
| 52 |
+
# New fields to add for each sentence
|
| 53 |
+
output_names:
|
| 54 |
+
begin: "response_begin"
|
| 55 |
+
end: "response_end"
|
| 56 |
+
text: "response_text"
|
| 57 |
+
- type: decode_sentences
|
| 58 |
+
source: "documents"
|
| 59 |
+
input_path: [~, "c"] # Null in path means wildcard
|
| 60 |
+
# New fields to add for each sentence
|
| 61 |
+
output_names:
|
| 62 |
+
document_id: "citation_doc_id"
|
| 63 |
+
begin: "citation_begin"
|
| 64 |
+
end: "citation_end"
|
| 65 |
+
text: "citation_text"
|
| 66 |
+
# Remove fields that we no longer need
|
| 67 |
+
- type: project
|
| 68 |
+
input_path: []
|
| 69 |
+
retained_fields:
|
| 70 |
+
- "response_begin"
|
| 71 |
+
- "response_end"
|
| 72 |
+
- "response_text"
|
| 73 |
+
- "citation_doc_id"
|
| 74 |
+
- "citation_begin"
|
| 75 |
+
- "citation_end"
|
| 76 |
+
- "citation_text"
|
| 77 |
+
# Merge adjacent document spans
|
| 78 |
+
- type: merge_spans
|
| 79 |
+
input_path: []
|
| 80 |
+
group_fields: ["response_begin", "response_end", "response_text", "citation_doc_id"]
|
| 81 |
+
begin_field: "citation_begin"
|
| 82 |
+
end_field: "citation_end"
|
| 83 |
+
text_field: "citation_text"
|
| 84 |
+
|
| 85 |
+
instruction: >
|
| 86 |
+
Split the last assistant response into individual sentences.
|
| 87 |
+
For each sentence in the response, identify the statement IDs from the below
|
| 88 |
+
documents that it references. Ensure that your output includes all response
|
| 89 |
+
sentence IDs, and for each response sentence ID, provide the list of corresponding
|
| 90 |
+
referring document sentence IDs. The output must be a json structure.
|
| 91 |
+
parameters:
|
| 92 |
+
temperature: 0.0
|
| 93 |
+
max_completion_tokens: 4096
|
| 94 |
+
sentence_boundaries:
|
| 95 |
+
# Mapping from string location to sentence delimiter prefix
|
| 96 |
+
last_message: "r" # <r0>, <r1>, etc.
|
| 97 |
+
documents: "c"
|
| 98 |
+
|
io_configs/context-attribution/io.yaml
ADDED
|
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: context-attribution
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"$defs": {
|
| 8 |
+
"_MODEL_OUTPUT_ENTRY": {
|
| 9 |
+
"properties": {
|
| 10 |
+
"r": {
|
| 11 |
+
"minimum": 0,
|
| 12 |
+
"title": "R",
|
| 13 |
+
"type": "integer"
|
| 14 |
+
},
|
| 15 |
+
"c": {
|
| 16 |
+
"items": {
|
| 17 |
+
"minimum": 0,
|
| 18 |
+
"type": "integer"
|
| 19 |
+
},
|
| 20 |
+
"title": "C",
|
| 21 |
+
"type": "array"
|
| 22 |
+
}
|
| 23 |
+
},
|
| 24 |
+
"required": [
|
| 25 |
+
"r",
|
| 26 |
+
"c"
|
| 27 |
+
],
|
| 28 |
+
"title": "_MODEL_OUTPUT_ENTRY",
|
| 29 |
+
"type": "object"
|
| 30 |
+
}
|
| 31 |
+
},
|
| 32 |
+
"items": {
|
| 33 |
+
"$ref": "#/$defs/_MODEL_OUTPUT_ENTRY"
|
| 34 |
+
},
|
| 35 |
+
"title": "_MODEL_OUTPUT",
|
| 36 |
+
"type": "array"
|
| 37 |
+
}
|
| 38 |
+
transformations:
|
| 39 |
+
# Explode lists of context sentence indices
|
| 40 |
+
- type: explode
|
| 41 |
+
input_path: [] # Zero-length path means match root element
|
| 42 |
+
target_field: "c"
|
| 43 |
+
# Model may repeat itself; drop the resulting duplicates.
|
| 44 |
+
- type: drop_duplicates
|
| 45 |
+
input_path: [] # Zero-length path means match root element
|
| 46 |
+
target_fields: ["r", "c"]
|
| 47 |
+
# Replace sentence number with sentence location and contents.
|
| 48 |
+
# Do this first for sentences from the last turn, then for sentences from context.
|
| 49 |
+
- type: decode_sentences
|
| 50 |
+
source: "last_message"
|
| 51 |
+
input_path: [~, "r"] # Null in path means wildcard
|
| 52 |
+
# New fields to add for each sentence
|
| 53 |
+
output_names:
|
| 54 |
+
begin: "response_begin"
|
| 55 |
+
end: "response_end"
|
| 56 |
+
text: "response_text"
|
| 57 |
+
- type: decode_sentences
|
| 58 |
+
source:
|
| 59 |
+
- "documents"
|
| 60 |
+
- "all_but_last_message"
|
| 61 |
+
input_path: [~, "c"] # Null in path means wildcard
|
| 62 |
+
# New fields to add for each sentence
|
| 63 |
+
output_names:
|
| 64 |
+
document_id: "attribution_doc_id"
|
| 65 |
+
message_index: "attribution_msg_index"
|
| 66 |
+
begin: "attribution_begin"
|
| 67 |
+
end: "attribution_end"
|
| 68 |
+
text: "attribution_text"
|
| 69 |
+
# Remove fields that we no longer need
|
| 70 |
+
- type: project
|
| 71 |
+
input_path: []
|
| 72 |
+
retained_fields:
|
| 73 |
+
- "response_begin"
|
| 74 |
+
- "response_end"
|
| 75 |
+
- "response_text"
|
| 76 |
+
- "attribution_doc_id"
|
| 77 |
+
- "attribution_msg_index"
|
| 78 |
+
- "attribution_begin"
|
| 79 |
+
- "attribution_end"
|
| 80 |
+
- "attribution_text"
|
| 81 |
+
|
| 82 |
+
instruction: >
|
| 83 |
+
You provided the last assistant response above based on context, which may include documents and/or previous conversation turns.
|
| 84 |
+
Your response is divided into sentences, numbered in the format <r0> sentence 0 <r1> sentence 1 ...
|
| 85 |
+
Sentences in the context are also numbered: <c0> sentence 0 <c1> sentence 1 ...
|
| 86 |
+
For each response sentence, please list the context sentences that were most important for you to generate the response sentence.
|
| 87 |
+
Provide your answer in JSON format, as an array of JSON objects, where each object has two members:
|
| 88 |
+
"r" with the response sentence number as the value, and "c" with an array of context sentence numbers as the value.
|
| 89 |
+
An example of such an array of objects is [{{"r": 0, "c": [3, 1, 4]}}, {{"r": 1, "c": [1, 5]}}].
|
| 90 |
+
List the context sentences in order from most important to least important.
|
| 91 |
+
Ensure that you include an object for each response sentence, even if the corresponding array of context sentence numbers is empty.
|
| 92 |
+
Answer with only the JSON and do not explain.
|
| 93 |
+
parameters:
|
| 94 |
+
max_completion_tokens: 4096
|
| 95 |
+
temperature: 0.0
|
| 96 |
+
sentence_boundaries:
|
| 97 |
+
# Mapping from string location to sentence delimiter prefix
|
| 98 |
+
last_message: "r" # <r0>, <r1>, etc.
|
| 99 |
+
documents: "c"
|
| 100 |
+
all_but_last_message: "c"
|
io_configs/context_relevance/io.yaml
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: context_relevance
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"title": "ContextRelevanceOutput",
|
| 8 |
+
"type": "object",
|
| 9 |
+
"properties": {
|
| 10 |
+
"context_relevance": {
|
| 11 |
+
"type": "string",
|
| 12 |
+
"description": "Context relevancy judgment.",
|
| 13 |
+
"enum": ["relevant", "irrelevant", "partially relevant"]
|
| 14 |
+
}
|
| 15 |
+
},
|
| 16 |
+
"required": ["context_relevance"]
|
| 17 |
+
}
|
| 18 |
+
# Additional turn of instructions to add to the chat
|
| 19 |
+
instruction: |
|
| 20 |
+
DOCUMENT: {document_content}
|
| 21 |
+
# Data transformations to perform during post-processing
|
| 22 |
+
transformations: ~
|
| 23 |
+
parameters:
|
| 24 |
+
temperature: 0.0
|
| 25 |
+
sentence_boundaries: ~
|
io_configs/factuality-correction/io.yaml
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Model name string, or null to use whatever is provided in the chat completion request.
|
| 2 |
+
name: factuality-correction
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"title": "Factuality Correction Intrinsic",
|
| 8 |
+
"type": "object",
|
| 9 |
+
"properties": {
|
| 10 |
+
"correction": {
|
| 11 |
+
"type": "string"
|
| 12 |
+
}
|
| 13 |
+
},
|
| 14 |
+
"required": ["correction"]
|
| 15 |
+
}
|
| 16 |
+
transformations: ~
|
| 17 |
+
instruction: ~
|
| 18 |
+
parameters:
|
| 19 |
+
# corrected response can several hundred tokens at high temperatures
|
| 20 |
+
max_completion_tokens: 4096
|
| 21 |
+
temperature: 0.0
|
| 22 |
+
# No sentence boundary detection
|
| 23 |
+
sentence_boundaries: ~
|
io_configs/factuality-detection/io.yaml
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Model name string, or null to use whatever is provided in the chat completion request.
|
| 2 |
+
name: factuality-detection
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"title": "Factuality Detection Intrinsic",
|
| 8 |
+
"type": "object",
|
| 9 |
+
"properties": {
|
| 10 |
+
"score": {
|
| 11 |
+
"type": "string",
|
| 12 |
+
"enum": ["yes", "no"]
|
| 13 |
+
}
|
| 14 |
+
},
|
| 15 |
+
"required": ["score"],
|
| 16 |
+
"additionalProperties": false
|
| 17 |
+
}
|
| 18 |
+
transformations: ~
|
| 19 |
+
instruction: ~
|
| 20 |
+
parameters:
|
| 21 |
+
max_completion_tokens: 20
|
| 22 |
+
temperature: 0.0
|
| 23 |
+
# No sentence boundary detection
|
| 24 |
+
sentence_boundaries: ~
|
io_configs/guardian-core/io.yaml
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: guardian-core
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"type": "object",
|
| 8 |
+
"properties": {
|
| 9 |
+
"score": {
|
| 10 |
+
"type": "string",
|
| 11 |
+
"enum": ["yes", "no"]
|
| 12 |
+
}
|
| 13 |
+
},
|
| 14 |
+
"required": ["score"],
|
| 15 |
+
"additionalProperties": false
|
| 16 |
+
}
|
| 17 |
+
# Output transformation rules to apply
|
| 18 |
+
transformations:
|
| 19 |
+
- type: likelihood
|
| 20 |
+
categories_to_values:
|
| 21 |
+
"yes": 1.0
|
| 22 |
+
"no": 0.0
|
| 23 |
+
input_path: ["score"]
|
| 24 |
+
# Convert scalar value to a record for consistency with other intrinsics
|
| 25 |
+
- type: nest
|
| 26 |
+
input_path: []
|
| 27 |
+
field_name: "guardian"
|
| 28 |
+
instruction: ~
|
| 29 |
+
parameters:
|
| 30 |
+
max_completion_tokens: 15
|
| 31 |
+
temperature: 0.0
|
| 32 |
+
sentence_boundaries: ~
|
io_configs/hallucination_detection/io.yaml
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: hallucination_detection
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"$defs": {
|
| 8 |
+
"HallucinationOutputEntry": {
|
| 9 |
+
"properties": {
|
| 10 |
+
"r": {
|
| 11 |
+
"minimum": 0,
|
| 12 |
+
"title": "Sentence Num",
|
| 13 |
+
"type": "integer"
|
| 14 |
+
},
|
| 15 |
+
"f": {
|
| 16 |
+
"title": "Is Faithful",
|
| 17 |
+
"type": "string",
|
| 18 |
+
"enum": ["faithful", "partial", "unfaithful", "NA"]
|
| 19 |
+
},
|
| 20 |
+
"e": {
|
| 21 |
+
"title": "Reasoning",
|
| 22 |
+
"type": "string"
|
| 23 |
+
}
|
| 24 |
+
},
|
| 25 |
+
"required": [
|
| 26 |
+
"r",
|
| 27 |
+
"e",
|
| 28 |
+
"f"
|
| 29 |
+
],
|
| 30 |
+
"title": "HallucinationOutputEntry",
|
| 31 |
+
"type": "object"
|
| 32 |
+
}
|
| 33 |
+
},
|
| 34 |
+
"items": {
|
| 35 |
+
"$ref": "#/$defs/HallucinationOutputEntry"
|
| 36 |
+
},
|
| 37 |
+
"title": "HallucinationOutput",
|
| 38 |
+
"type": "array"
|
| 39 |
+
}
|
| 40 |
+
transformations:
|
| 41 |
+
# Replace sentence number with sentence location and contents
|
| 42 |
+
- type: decode_sentences
|
| 43 |
+
source: "last_message"
|
| 44 |
+
input_path: [~, "r"] # Null in path means wildcard
|
| 45 |
+
# New fields to add for each sentence
|
| 46 |
+
output_names:
|
| 47 |
+
begin: "response_begin"
|
| 48 |
+
end: "response_end"
|
| 49 |
+
text: "response_text"
|
| 50 |
+
# Remove fields that we no longer need and rename some of the fields.
|
| 51 |
+
- type: project
|
| 52 |
+
input_path: []
|
| 53 |
+
retained_fields:
|
| 54 |
+
"response_begin": "response_begin"
|
| 55 |
+
"response_end": "response_end"
|
| 56 |
+
"response_text": "response_text"
|
| 57 |
+
"f": "faithfulness"
|
| 58 |
+
"e": "explanation"
|
| 59 |
+
instruction: >
|
| 60 |
+
Split the last assistant response into individual sentences.
|
| 61 |
+
For each sentence in the last assistant response, identify the faithfulness
|
| 62 |
+
by comparing with the provided documents and generate the faithfulness reasoning
|
| 63 |
+
and faithfulness decision.
|
| 64 |
+
Ensure that your output includes all response sentence IDs,
|
| 65 |
+
and for each response sentence ID, provide the corresponding faithfulness
|
| 66 |
+
reasoning and faithfulness decision.
|
| 67 |
+
The output must be a json structure.
|
| 68 |
+
parameters:
|
| 69 |
+
# Current LoRA can be quite verbose in its explanations.
|
| 70 |
+
max_completion_tokens: 4096
|
| 71 |
+
temperature: 0.0
|
| 72 |
+
sentence_boundaries:
|
| 73 |
+
last_message: "i"
|
io_configs/policy-guardrails/io.yaml
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: policy-guardrails
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"title": "Policy Guardrails Intrinsic",
|
| 8 |
+
"type": "object",
|
| 9 |
+
"properties": {
|
| 10 |
+
"score": {
|
| 11 |
+
"type": "string",
|
| 12 |
+
"enum": ["Yes", "No","Ambiguous"]
|
| 13 |
+
}
|
| 14 |
+
},
|
| 15 |
+
"required": ["score"],
|
| 16 |
+
"additionalProperties": false
|
| 17 |
+
}
|
| 18 |
+
transformations: ~
|
| 19 |
+
instruction: ~
|
| 20 |
+
parameters:
|
| 21 |
+
temperature: 0.0
|
| 22 |
+
max_completion_tokens: 20
|
| 23 |
+
# No sentence boundary detection
|
| 24 |
+
sentence_boundaries: ~
|
io_configs/query_clarification/io.yaml
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: query_clarification
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"properties": {
|
| 8 |
+
"clarification": {
|
| 9 |
+
"title": "clarification",
|
| 10 |
+
"type": "string"
|
| 11 |
+
}
|
| 12 |
+
},
|
| 13 |
+
"required": [
|
| 14 |
+
"clarification"
|
| 15 |
+
],
|
| 16 |
+
"title": "QueryClarificationOutput",
|
| 17 |
+
"type": "object"
|
| 18 |
+
}
|
| 19 |
+
transformations: ~
|
| 20 |
+
instruction: ~
|
| 21 |
+
parameters:
|
| 22 |
+
max_completion_tokens: 1024
|
| 23 |
+
temperature: 0.0
|
| 24 |
+
sentence_boundaries: false
|
io_configs/query_rewrite/io.yaml
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: query_rewrite
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"properties": {
|
| 8 |
+
"rewritten_question": {
|
| 9 |
+
"title": "Rewritten Question",
|
| 10 |
+
"type": "string"
|
| 11 |
+
}
|
| 12 |
+
},
|
| 13 |
+
"required": [
|
| 14 |
+
"rewritten_question"
|
| 15 |
+
],
|
| 16 |
+
"title": "QueryRewriteOutput",
|
| 17 |
+
"type": "object"
|
| 18 |
+
}
|
| 19 |
+
transformations: ~
|
| 20 |
+
instruction: ~
|
| 21 |
+
parameters:
|
| 22 |
+
max_completion_tokens: 1024
|
| 23 |
+
temperature: 0.0
|
| 24 |
+
sentence_boundaries: false
|
io_configs/requirement-check/io.yaml
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: requirement-check
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"type": "object",
|
| 8 |
+
"properties": {
|
| 9 |
+
"score": {
|
| 10 |
+
"type": "string",
|
| 11 |
+
"enum": ["yes", "no"]
|
| 12 |
+
}
|
| 13 |
+
},
|
| 14 |
+
"required": ["score"],
|
| 15 |
+
"additionalProperties": false
|
| 16 |
+
}
|
| 17 |
+
# Output transformation rules to apply
|
| 18 |
+
transformations:
|
| 19 |
+
- type: likelihood
|
| 20 |
+
categories_to_values:
|
| 21 |
+
"yes": 1.0
|
| 22 |
+
"no": 0.0
|
| 23 |
+
input_path: ["score"]
|
| 24 |
+
# Convert scalar value to a record for consistency with other intrinsics
|
| 25 |
+
- type: nest
|
| 26 |
+
input_path: []
|
| 27 |
+
field_name: "requirement_check"
|
| 28 |
+
instruction: ~
|
| 29 |
+
parameters:
|
| 30 |
+
max_completion_tokens: 15
|
| 31 |
+
temperature: 0.0
|
| 32 |
+
sentence_boundaries: ~
|
io_configs/uncertainty/io.yaml
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name: uncertainty
|
| 2 |
+
# Model name string, or null to use whatever is provided in the chat completion request
|
| 3 |
+
model: ~
|
| 4 |
+
# JSON schema of the model's output
|
| 5 |
+
response_format: |
|
| 6 |
+
{
|
| 7 |
+
"type": "object",
|
| 8 |
+
"properties": {
|
| 9 |
+
"score": {
|
| 10 |
+
"type": "string",
|
| 11 |
+
"enum": ["0", "1", "2", "3", "4", "5", "6", "7", "8", "9"]
|
| 12 |
+
}
|
| 13 |
+
},
|
| 14 |
+
"required": ["score"],
|
| 15 |
+
"additionalProperties": false
|
| 16 |
+
}
|
| 17 |
+
# Output transformation rules to apply
|
| 18 |
+
transformations:
|
| 19 |
+
- type: likelihood
|
| 20 |
+
categories_to_values:
|
| 21 |
+
# Each 1-digit output maps to 0.1 * <output> + 0.05
|
| 22 |
+
"0": 0.05
|
| 23 |
+
"1": 0.15
|
| 24 |
+
"2": 0.25
|
| 25 |
+
"3": 0.35
|
| 26 |
+
"4": 0.45
|
| 27 |
+
"5": 0.55
|
| 28 |
+
"6": 0.65
|
| 29 |
+
"7": 0.75
|
| 30 |
+
"8": 0.85
|
| 31 |
+
"9": 0.95
|
| 32 |
+
input_path: ["score"]
|
| 33 |
+
# Convert scalar value to a record for consistency with other intrinsics
|
| 34 |
+
- type: project
|
| 35 |
+
input_path: []
|
| 36 |
+
retained_fields:
|
| 37 |
+
score: "certainty"
|
| 38 |
+
instruction: ~
|
| 39 |
+
parameters:
|
| 40 |
+
max_completion_tokens: 15
|
| 41 |
+
temperature: 0.0
|
| 42 |
+
sentence_boundaries: ~
|
merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a3f64677efd51cba02b9f473401798d7286517f4d1c20b400a18fb5df39f054e
|
| 3 |
+
size 10043422843
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": {
|
| 3 |
+
"content": "<|end_of_text|>",
|
| 4 |
+
"lstrip": false,
|
| 5 |
+
"normalized": false,
|
| 6 |
+
"rstrip": false,
|
| 7 |
+
"single_word": false
|
| 8 |
+
},
|
| 9 |
+
"eos_token": {
|
| 10 |
+
"content": "<|end_of_text|>",
|
| 11 |
+
"lstrip": false,
|
| 12 |
+
"normalized": false,
|
| 13 |
+
"rstrip": false,
|
| 14 |
+
"single_word": false
|
| 15 |
+
},
|
| 16 |
+
"pad_token": {
|
| 17 |
+
"content": "<|pad|>",
|
| 18 |
+
"lstrip": false,
|
| 19 |
+
"normalized": false,
|
| 20 |
+
"rstrip": false,
|
| 21 |
+
"single_word": false
|
| 22 |
+
},
|
| 23 |
+
"unk_token": {
|
| 24 |
+
"content": "<|unk|>",
|
| 25 |
+
"lstrip": false,
|
| 26 |
+
"normalized": false,
|
| 27 |
+
"rstrip": false,
|
| 28 |
+
"single_word": false
|
| 29 |
+
}
|
| 30 |
+
}
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"backend": "tokenizers",
|
| 4 |
+
"bos_token": "<|end_of_text|>",
|
| 5 |
+
"clean_up_tokenization_spaces": false,
|
| 6 |
+
"eos_token": "<|end_of_text|>",
|
| 7 |
+
"errors": "replace",
|
| 8 |
+
"extra_special_tokens": [
|
| 9 |
+
"<|citations|>",
|
| 10 |
+
"<|query_rewrite|>",
|
| 11 |
+
"<|query_clarification|>",
|
| 12 |
+
"<|context_relevance|>",
|
| 13 |
+
"<|hallucination_detection|>",
|
| 14 |
+
"<|answerability|>",
|
| 15 |
+
"<|uncertainty|>",
|
| 16 |
+
"<|requirement-check|>",
|
| 17 |
+
"<|context-attribution|>",
|
| 18 |
+
"<|factuality-detection|>",
|
| 19 |
+
"<|policy-guardrails|>",
|
| 20 |
+
"<|factuality-correction|>",
|
| 21 |
+
"<|guardian-core|>"
|
| 22 |
+
],
|
| 23 |
+
"is_local": true,
|
| 24 |
+
"local_files_only": false,
|
| 25 |
+
"model_max_length": 1000000000000000019884624838656,
|
| 26 |
+
"pad_token": "<|pad|>",
|
| 27 |
+
"padding_side": "left",
|
| 28 |
+
"tokenizer_class": "GPT2Tokenizer",
|
| 29 |
+
"unk_token": "<|unk|>"
|
| 30 |
+
}
|
vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|