This is the first iteration of fine-tuning MamayLM for function calling. In the future, we plan to add metrics and improve training.
During this phase new tokens (including tool_call) were introduced to the model and we evaluated how well it uses and understands the purpose of tool_call.

Metrics

Accuracy in function calling (if response contains tool_call token) - find_longest_common_sequence_length(ground_truth_tokens, generated_tokens) / len(ground_truth_tokens)
Match in helpful exchange (if response does not contain tool_call token) - Computes the percentage of matching elements between generated tokens and ground truth tokens

Performance before fine-tuning:

Accuracy in function calling: 0.38107
Match in helpful exchange: 0.07440

Performance after fine-tuning:

Accuracy in function calling: 0.95415
Match in helpful exchange: 0.09937

Quick start

python

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

peft_model_id = "TymofiiNasobko/MamayLM-function-calling"
peftconfig = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(
    peftconfig.base_model_name_or_path,
    attn_implementation="eager",
    device_map=device,
)
tokenizer = AutoTokenizer.from_pretrained(peft_model_id)
model.resize_token_embeddings(len(tokenizer))
model = PeftModel.from_pretrained(model, peft_model_id)
model = model.to(compute_dtype)
model = model.eval()

Training procedure

This model was trained with SFT.

Framework versions

TRL: 0.25.0
Transformers: 4.57.1
Pytorch: 2.8.0+cu128
Datasets: 4.4.1
Tokenizers: 0.22.1

Citations

Cite TRL as:

bibtex

@misc{vonwerra2022trl,
	title        = {{TRL: Transformer Reinforcement Learning}},
	author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
	year         = 2020,
	journal      = {GitHub repository},
	publisher    = {GitHub},
	howpublished = {\url{https://github.com/huggingface/trl}}
}