Issue with finetuning Gemma 3 with "train_on_responses_only"

[MostAardvark224]

Hey all, I'm new to unsloth and was wondering if anyone could help me solve an issue with finetuning Gemma 3.

Here's my code: (for context most of this is from the unsloth colab notebook on finetuning Gemma 3, I just adapted it for my own dataset).

# Loading the model
model, tokenizer = FastModel.from_pretrained(
    model_name = "unsloth/gemma-3-4b-it",
    max_seq_length = 2048,
    load_in_4bit = True,  
    load_in_8bit = False, 
    full_finetuning = False
)
model = FastModel.get_peft_model(
    model,
    finetune_vision_layers     = False, 
    finetune_language_layers   = True,  
    finetune_attention_modules = True, 
    finetune_mlp_modules       = True,  
    r = 8,          
    lora_alpha = 8,  
    lora_dropout = 0,
    bias = "none",
    random_state = 3407,
)
from unsloth.chat_templates import get_chat_template
tokenizer = get_chat_template(
    tokenizer,
    chat_template = "gemma-3",
)
from datasets import load_dataset
dataset = load_dataset("MostAardvark224/mydataset", split = "train") # This is my own private dataset I'm trying to finetune on. It has two columns: "prompt" and "completion".
from unsloth.chat_templates import standardize_data_formats
dataset = standardize_data_formats(dataset)
def to_conversations(batch): # This function converts my two column dataset into a single column "conversations".
    return {
        "conversations": [
            [
                {"role": "user",  "content": p},
                {"role": "model", "content": c},
            ]
            for p, c in zip(batch["prompt"], batch["completion"])
        ]
    }

dataset = dataset.map(to_conversations, batched=True, remove_columns=["prompt", "completion"])
def formatting_prompts_func(examples): # formatting func that was given in the notebook
   convos = examples["conversations"]
   texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False).removeprefix('<bos>') for convo in convos]
   return { "text" : texts, }
dataset = dataset.map(formatting_prompts_func, batched = True)
dataset[0]["text"]

When I print out the row, this is what it looks like:

'<start_of_turn>user\n my prompt xyz <end_of_turn>\n<start_of_turn>model\n{"model completion as JSON object"}<end_of_turn>\n'
which is what I think the Gemma 3 chat template is supposed to look like (it's just missing the token.

I then initialize my SFTTrainer

from trl import SFTTrainer, SFTConfig
trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    eval_dataset = None, # Can set up evaluation!
    args = args
Finally, I attempt to train on responses only, but this is where I get hit with an error.

from unsloth.chat_templates import train_on_responses_only
trainer = train_on_responses_only(
    trainer,
    instruction_part = "<start_of_turn>user\n",
    response_part = "<start_of_turn>model\n",
)

Error:

---------------------------------------------------------------------------
ZeroDivisionError                         Traceback (most recent call last)
/tmp/ipykernel_228/697443393.py in <cell line: 0>()
      1 from unsloth.chat_templates import train_on_responses_only
----> 2 trainer = train_on_responses_only(
      3     trainer,
      4     instruction_part = "<start_of_turn>user\n",
      5     response_part = "<start_of_turn>model\n",

/usr/local/lib/python3.11/dist-packages/unsloth_zoo/dataset_utils.py in train_on_responses_only(trainer, instruction_part, response_part, force_match, tokenizer, return_function, num_proc)
    369     # Check if all labels randomnly got masked to nothing - maybe wrong chat template?
    370     from .training_utils import fix_zero_training_loss
--> 371     fix_zero_training_loss(None, tokenizer, trainer.train_dataset)
    372     return trainer
    373 pass

/usr/local/lib/python3.11/dist-packages/torch/utils/_contextlib.py in decorate_context(*args, **kwargs)
    114     def decorate_context(*args, **kwargs):
    115         with ctx_factory():
--> 116             return func(*args, **kwargs)
    117 
    118     return decorate_context

/usr/local/lib/python3.11/dist-packages/unsloth_zoo/training_utils.py in fix_zero_training_loss(model, tokenizer, train_dataset)
     70 
     71         elif seen_bad / (seen_bad + seen_good) == 1:
---> 72             raise ZeroDivisionError(
     73                 "Unsloth: All labels in your dataset are -100. Training losses will be all 0.\n"\
     74                 "For example, are you sure you used `train_on_responses_only` correctly?\n"\

ZeroDivisionError: Unsloth: All labels in your dataset are -100. Training losses will be all 0.
For example, are you sure you used `train_on_responses_only` correctly?
Or did you mask our tokens incorrectly? Maybe this is intended?
Maybe you're using a Llama chat template on a non Llama model for example?

I've looked all around and can't really find any solutions. I think the issue likely has something to do with my dataset because if I use the "Finetome-100k" dataset that was used in the original notebook it works just fine. I just can't pinpoint where the error is coming from exactly.

Any help would be MUCH appreciated. Please ask further questions if more specifics are required.

[MostAardvark224]

Fix (kind of): I used my same dataset on Qwen3-32B and it worked just fine, so if you're having the same issue try switching models.