Funcdex-1.7B

Funcdex-1.7B

<div align="center"> <img src="assets/funcdex_hero.png" alt="Funcdex Hero" width="70%"> </div>

Funcdex-1.7B is a research preview model by Prem Labs. It has been trained on a mix of Funcdex-MT-Function-Calling, Instruct-Following, Single-turn function datasets. It is a LoRA finetune of Qwen3-1.7B (with thinking disabled).

This model excels at Multi-turn Function Calling with tools from gmail, jira, calendar, docs, etc.

The code used to generate the dataset can be found here.

Evaluation

<div align="center"> <img src="assets/line_plot.png" alt="Line Plot" width="80%"> </div>

Notes:

• Funcdex-0.6B is the average of performances of individual Funcdex-0.6B models.
• For cost, we track the number of prompt/completion tokens for evaluating 300 conversations.
• e.g. If token cost is input=$1 and output=$10 per million tokens, and evaluation needed 0.5M and 0.1M input/output tokens, then cost is 1 * 0.5 + 0.1 * 10 = $1.5.
• Qwen3-0.6B and Qwen3-1.7B evaluation costs are estimated by extrapolating from Llama3.2-3B serverless costs. Other model's costs are sourced from Openrouter.

Results

BFCL v3

• We filtered BFCLv3 examples relevant to the toolkits/bundles and report performance:
• The filtered set is only 83 examples. Further emphasizing the need for workflow/toolkit-specialized workflows.

<table border="1" class="dataframe"> <thead> <tr style="text-align: center;"> <th>LLM</th> <th>Acc %</th> </tr> </thead> <tbody> <tr style="text-align: center;"> <td>GPT-5 Mini<br>(medium)</td> <td>0.71</td> </tr> <tr style="text-align: center;"> <td>Qwen3-1.7B</td> <td>0.82</td> </tr> <tr style="text-align: center;"> <td><strong><a href="https://huggingface.co/prem-research/Funcdex-1.7B">Funcdex-1.7B</a><strong></td> <td><strong>0.86</strong></td> </tr> </tbody> </table>

Funcdex-MT: Overall Performance

<table border="1" class="dataframe"> <thead> <tr style="text-align: center;"> <th>LLM</th> <th>Exact Match</th> <th>String Ratio</th> <th>Total Cost ($)</th> </tr> </thead> <tbody> <tr style="text-align: center;"> <td>GPT-OSS-120B<br>(medium)</td> <td>0.35</td> <td>0.51</td> <td>9.32</td> </tr> <tr style="text-align: center;"> <td>GPT-5 Mini<br>(medium)</td> <td>0.35</td> <td>0.58</td> <td>99.71</td> </tr> <tr style="text-align: center;"> <td>GPT-5<br>(minimal)</td> <td>0.18</td> <td>0.59</td> <td>205.45</td> </tr> <tr style="text-align: center;"> <td>Qwen3-0.6B</td> <td>0.27</td> <td>0.59</td> <td>2.83</td> </tr> <tr style="text-align: center;"> <td>Qwen3-1.7B</td> <td>0.27</td> <td>0.69</td> <td>5.73</td> </tr> <tr style="text-align: center;"> <td><strong><a href="https://huggingface.co/collections/prem-research/funcdex">Funcdex-0.6B</a></strong></td> <td><strong>0.39</strong></td> <td><strong>0.70</strong></td> <td><strong>0.19</strong></td> </tr> <tr style="text-align: center;"> <td><strong><a href="https://huggingface.co/prem-research/Funcdex-1.7B">Funcdex-1.7B</a></strong></td> <td><strong>0.43</strong></td> <td><strong>0.81</strong></td> <td>5.64</td> </tr> </tbody> </table>

Funcdex-MT: Toolkit-Level Performance

<table border="1" class="dataframe"> <thead> <tr style="text-align: center;"> <th rowspan="2">Toolkit</th> <th colspan="2">GPT-OSS-120B<br>(medium)</th> <th colspan="2">GPT-5<br>(minimal)</th> <th colspan="2">GPT-5 Mini<br>(medium)</th> <th colspan="2">Qwen3-0.6B</th> <th colspan="3">Funcdex-0.6B</th> <th colspan="2">Qwen3-1.7B</th> <th colspan="3">Funcdex-1.7B</th> </tr> <tr style="text-align: center;"> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>LoRA Checkpoint</th> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>LoRA Checkpoint</th> </tr> </thead> <tbody> <tr style="text-align: center;"> <td><img src="assets/icons/asana.png" width="20" height="20" style="vertical-align: middle;"/> Asana</td> <td>0.38</td> <td>0.47</td> <td>0.12</td> <td>0.68</td> <td>0.49</td> <td>0.71</td> <td>0.33</td> <td>0.63</td> <td>0.46</td> <td>0.69</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-asana">🤗</a></td> <td>0.30</td> <td>0.79</td> <td>0.52</td> <td>0.82</td> <td rowspan="10"><a href="https://huggingface.co/prem-research/Funcdex-1.7B">🤗</a></td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/calendly.png" width="20" height="20" style="vertical-align: middle;"/> Calendly</td> <td>0.47</td> <td>0.56</td> <td>0.41</td> <td>0.63</td> <td>0.41</td> <td>0.56</td> <td>0.44</td> <td>0.66</td> <td>0.54</td> <td>0.78</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-calendly">🤗</a></td> <td>0.47</td> <td>0.74</td> <td>0.54</td> <td>0.86</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/gmail.png" width="20" height="20" style="vertical-align: middle;"/> Gmail</td> <td>0.48</td> <td>0.70</td> <td>0.24</td> <td>0.69</td> <td>0.50</td> <td>0.73</td> <td>0.27</td> <td>0.61</td> <td>0.47</td> <td>0.72</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-gmail">🤗</a></td> <td>0.31</td> <td>0.73</td> <td>0.53</td> <td>0.83</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/google-calendar.png" width="20" height="20" style="vertical-align: middle;"/> Calendar</td> <td>0.27</td> <td>0.52</td> <td>0.20</td> <td>0.50</td> <td>0.21</td> <td>0.51</td> <td>0.21</td> <td>0.53</td> <td>0.39</td> <td>0.74</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-googlecalendar">🤗</a></td> <td>0.23</td> <td>0.64</td> <td>0.47</td> <td>0.83</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/docs.png" width="20" height="20" style="vertical-align: middle;"/> Docs</td> <td>0.19</td> <td>0.38</td> <td>0.07</td> <td>0.49</td> <td>0.18</td> <td>0.46</td> <td>0.07</td> <td>0.58</td> <td>0.13</td> <td>0.64</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-googledocs">🤗</a></td> <td>0.11</td> <td>0.62</td> <td>0.18</td> <td>0.79</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/google-drive.png" width="20" height="20" style="vertical-align: middle;"/> Drive</td> <td>0.34</td> <td>0.52</td> <td>0.19</td> <td>0.61</td> <td>0.38</td> <td>0.58</td> <td>0.26</td> <td>0.65</td> <td>0.40</td> <td>0.75</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-googledrive">🤗</a></td> <td>0.26</td> <td>0.73</td> <td>0.48</td> <td>0.82</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/jira.png" width="20" height="20" style="vertical-align: middle;"/> Jira</td> <td>0.47</td> <td>0.53</td> <td>0.17</td> <td>0.65</td> <td>0.47</td> <td>0.66</td> <td>0.51</td> <td>0.69</td> <td>0.58</td> <td>0.76</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-jira">🤗</a></td> <td>0.47</td> <td>0.76</td> <td>0.59</td> <td>0.83</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/stripe.png" width="20" height="20" style="vertical-align: middle;"/> Stripe</td> <td>0.15</td> <td>0.37</td> <td>0.10</td> <td>0.46</td> <td>0.12</td> <td>0.39</td> <td>0.08</td> <td>0.50</td> <td>0.17</td> <td>0.71</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-stripe">🤗</a></td> <td>0.09</td> <td>0.56</td> <td>0.16</td> <td>0.80</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/to-do-list.png" width="20" height="20" style="vertical-align: middle;"/> Todoist</td> <td>0.65</td> <td>0.74</td> <td>0.19</td> <td>0.72</td> <td>0.64</td> <td>0.79</td> <td>0.57</td> <td>0.87</td> <td>0.65</td> <td>0.88</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-todoist">🤗</a></td> <td>0.55</td> <td>0.91</td> <td>0.72</td> <td>0.94</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/whatsapp.png" width="20" height="20" style="vertical-align: middle;"/> Whatsapp</td> <td>0.23</td> <td>0.39</td> <td>0.13</td> <td>0.47</td> <td>0.24</td> <td>0.43</td> <td>0.20</td> <td>0.43</td> <td>0.28</td> <td>0.64</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-whatsapp">🤗</a></td> <td>0.26</td> <td>0.55</td> <td>0.31</td> <td>0.71</td> </tr> </tbody> </table>

• Funcdex-0.6B are specialized models. Reported number is the average performance of each specific model in their respective subset.

Funcdex-MT: Bundle/Multi-toolkit Performance:

<table border="1" class="dataframe"> <thead> <tr style="text-align: center;"> <th rowspan="2">Bundle</th> <th colspan="2">GPT-OSS-120B<br>(medium)</th> <th colspan="2">GPT-5<br>(minimal)</th> <th colspan="2">GPT-5 Mini<br>(medium)</th> <th colspan="2">Qwen3-0.6B</th> <th colspan="3">Funcdex-0.6B</th> <th colspan="2">Qwen3-1.7B</th> <th colspan="3">Funcdex-1.7B</th> </tr> <tr style="text-align: center;"> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>LoRA Checkpoint</th> <th>EM</th> <th>SR</th> <th>EM</th> <th>SR</th> <th>LoRA Checkpoint</th> </tr> </thead> <tbody> <tr style="text-align: center;"> <td><img src="assets/icons/gmail.png" width="20" height="20" style="vertical-align: middle;"/>Gmail<img src="assets/icons/google-calendar.png" width="20" height="20" style="vertical-align: middle;"/>Calendar</td> <td>0.28</td> <td>0.53</td> <td>0.15</td> <td>0.54</td> <td>0.22</td> <td>0.56</td> <td>0.19</td> <td>0.51</td> <td>0.26</td> <td>0.54</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-gmail_googlecalendar">🤗</a></td> <td>0.17</td> <td>0.61</td> <td>0.32</td> <td>0.71</td> <td rowspan="5"><a href="https://huggingface.co/prem-research/Funcdex-1.7B">🤗</a></td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/google-drive.png" width="20" height="20" style="vertical-align: middle;"/>Drive <img src="assets/icons/calendly.png" width="20" height="20" style="vertical-align: middle;"/> Calendly <img src="assets/icons/google-calendar.png" width="20" height="20" style="vertical-align: middle;"/> Calendar</td> <td>0.32</td> <td>0.45</td> <td>0.17</td> <td>0.52</td> <td>0.35</td> <td>0.47</td> <td>0.19</td> <td>0.49</td> <td>0.35</td> <td>0.60</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-googledrive_calendly_googlecalendar">🤗</a></td> <td>0.15</td> <td>0.66</td> <td>0.40</td> <td>0.78</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/google-drive.png" width="20" height="20" style="vertical-align: middle;"/>Drive <img src="assets/icons/docs.png" width="20" height="20" style="vertical-align: middle;"/> Docs</td> <td>0.28</td> <td>0.37</td> <td>0.12</td> <td>0.50</td> <td>0.33</td> <td>0.47</td> <td>0.18</td> <td>0.54</td> <td>0.34</td> <td>0.70</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-googledrive_googledocs">🤗</a></td> <td>0.19</td> <td>0.68</td> <td>0.43</td> <td>0.76</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/jira.png" width="20" height="20" style="vertical-align: middle;"/>Jira <img src="assets/icons/gmail.png" width="20" height="20" style="vertical-align: middle;"/> Gmail</td> <td>0.42</td> <td>0.60</td> <td>0.18</td> <td>0.66</td> <td>0.36</td> <td>0.66</td> <td>0.29</td> <td>0.61</td> <td>0.39</td> <td>0.71</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-jira_gmail">🤗</a></td> <td>0.28</td> <td>0.72</td> <td>0.44</td> <td>0.82</td> </tr> <tr style="text-align: center;"> <td><img src="assets/icons/whatsapp.png" width="20" height="20" style="vertical-align: middle;"/>Whatsapp <img src="assets/icons/to-do-list.png" width="20" height="20" style="vertical-align: middle;"/> Todoist</td> <td>0.32</td> <td>0.58</td> <td>0.19</td> <td>0.66</td> <td>0.35</td> <td>0.69</td> <td>0.26</td> <td>0.50</td> <td>0.41</td> <td>0.70</td> <td><a href="https://huggingface.co/prem-research/Funcdex-0.6B-whatsapp_todoist">🤗</a></td> <td>0.27</td> <td>0.68</td> <td>0.39</td> <td>0.77</td> </tr> </tbody> </table>

Inference

• Given a conversation, we extract all tuples (context_messages, function_calls) and use it to generate predictions. We ignore the content field and only evaluate function_calls generated by an LLM.
• We use vLLM deployment with tool_choice="auto".

Metrics

Given a list of predicted and reference function calls, we report two metrics:

• Function Call String Match (SR): We perform greedy match and report best-matched string ratio using difflib.SequenceMatcher.ratio. The number reported is average string ratio.
• Exact Match (EM): Same as above, but we perform exact string match instead. The number reported is EM F1 Score.

EM is a strict metric, and penalizes string arguments in function calls that may be "okay", e.g. "email_content": "This is an example." v/s "email_content": "This is an Example.", both only differ by one letter.

Deployment with vLLM

vllm serve ojus1/Qwen3-1.7B-Instruct --enable-lora --lora-modules prem-research/Funcdex-1.7B=prem-research/Funcdex-1.7B --enable-auto-tool-choice --tool-call-parser hermes

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
import json

# Load model and tokenizer
base_model_name = "ojus1/Qwen3-1.7B-Instruct"
model_name = "prem-research/Funcdex-1.7B"

tokenizer = AutoTokenizer.from_pretrained(model_name)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype="auto",
    device_map="auto"
)

model = PeftModel.from_pretrained(
    base_model,
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# Define tools (supports all toolkits)
tools = [
    {
        "type": "function",
        "function": {
            "name": "CREATE_SHARED_DRIVE",
            "description": "Create a new shared drive in Google Drive",
            "parameters": {
                "type": "object",
                "properties": {
                    "name": {"type": "string", "description": "Name of the shared drive"},
                    "requestId": {"type": "string", "description": "Unique request ID"}
                },
                "required": ["name", "requestId"]
            }
        }
    },
    {
        "type": "function",
        "function": {
            "name": "CREATE_A_FOLDER",
            "description": "Create a folder in Google Drive",
            "parameters": {
                "type": "object",
                "properties": {
                    "folder_name": {"type": "string", "description": "Name of the folder"},
                    "parent_id": {"type": "string", "description": "Parent drive or folder ID"}
                },
                "required": ["folder_name", "parent_id"]
            }
        }
    }
]

# Define conversation
messages = [
    {"role": "system", "content": "You are a helpful assistant that can help with tasks by using tools."},
    {"role": "user", "content": "Create a shared drive named 'Partner-Alpha-Integration' with request ID 'req-12345'."}
]

# Apply chat template with tools
formatted_input = tokenizer.apply_chat_template(
    messages,
    tools=tools,
    tokenize=False,
    add_generation_prompt=True
)

# Tokenize and generate
input_tokens = tokenizer(formatted_input, return_tensors="pt").to(model.device)
output = model.generate(**input_tokens, max_new_tokens=256, do_sample=False)
response = tokenizer.decode(output[0][input_tokens['input_ids'].shape[1]:], skip_special_tokens=True)

print("Response:", response)
# Expected output includes: <tool_call>{"name": "CREATE_SHARED_DRIVE", "arguments": {"name": "Partner-Alpha-Integration", "requestId": "req-12345"}}</tool_call>

For best results, provide detailed system-prompt to steer the tool-use behaviour.

License

The models, code and the dataset are licensed under MIT License.