Finetune fastchat with Zephyr format

[nghidinhit]

How can I perform fine-tuning on FastChat using the Zephyr format? I've noticed that within the preprocess function, there is hardcoded logic intended for fine-tuning with the Vicuna template.

[jwong8314]

You can just change the "vicuna" to "zephyr". The bigger issue is the rest of preprocess where you need to rewrite the hardcoded mask generation.

I wrote something like the following to support dolphin's format:

 if dolphin_format:
            turns = conversation.split(sep)

            cur_len = 1
            target[:cur_len] = IGNORE_TOKEN_ID
            for i, turn in enumerate(turns):
                if turn == "":
                    break
                turn_len = len(tokenizer(turn).input_ids)           
                
                if turn.split("\n")[0] != conv.roles[1]:
                    target[cur_len: cur_len + turn_len + 2] = IGNORE_TOKEN_ID
                elif  turn.split("\n")[0] == conv.roles[1]:
                    size_of_role = len(tokenizer(conv.roles[1]).input_ids)
                    target[cur_len: cur_len + size_of_role]  = IGNORE_TOKEN_ID
                    target[cur_len + turn_len: cur_len+turn_len+2]  = IGNORE_TOKEN_ID

                cur_len += turn_len + 2
            target[cur_len:] = IGNORE_TOKEN_ID

If I have some time later, I'll push a PR that should make finetuning agnostic to formatting. In the meantime, feel free to just switch to dolphin's format.

[nghidinhit]

This is my custom code:
`
def preprocess(
sources,
tokenizer: transformers.PreTrainedTokenizer,
) -> Dict:
conv = get_conversation_template("zephyr")

roles = {"human": conv.roles[0], "gpt": conv.roles[1]}

# Apply prompt templates
conversations = []
for i, source in enumerate(sources):
    conv.system_message = source[0]["value"].strip()
    source = source[1:]

    conv.messages = []
    for j, sentence in enumerate(source):
        role = roles[sentence["from"]]
        assert role == conv.roles[j % 2], f"{i}"
        conv.append_message(role, sentence["value"])
    conversations.append(conv.get_prompt())

# Tokenize conversations
input_ids = tokenizer(
    conversations,
    return_tensors="pt",
    padding="max_length",
    max_length=tokenizer.model_max_length,
    truncation=True,
).input_ids

targets = input_ids.clone()

# Mask targets. Only compute loss on the assistant outputs.
sep = conv.roles[1] + "\n"
for conversation, target in zip(conversations, targets):
    total_len = int(target.ne(tokenizer.pad_token_id).sum())

    turns = conversation.split("<|user|>\n")
    cur_len = 1  # for <s> special character
    target[:cur_len] = IGNORE_TOKEN_ID

    for i, turn in enumerate(turns):
        if turn == "":
            break

        if i == 0:  # system message
            parts = [turn, ""]
        else:
            turn = f"<|user|>\n{turn}"
            parts = turn.split(sep)  # user text and assistant text
            if len(parts) != 2:
                break
            parts[0] += sep
        turn_len = len(tokenizer(turn).input_ids) - 1  # loại bỏ kí tự <s>
        instruction_len = len(tokenizer(parts[0]).input_ids) - 1
        target[cur_len: cur_len + instruction_len] = IGNORE_TOKEN_ID

        cur_len += turn_len

    target[cur_len:] = IGNORE_TOKEN_ID

    if False:  # Inspect and check the correctness of masking
        z = target.clone()
        z = torch.where(z == IGNORE_TOKEN_ID, tokenizer.unk_token_id, z)
        rank0_print(tokenizer.decode(z))
        exit()

    if cur_len < tokenizer.model_max_length:
        if cur_len != total_len:
            target[:] = IGNORE_TOKEN_ID
            rank0_print(
                f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
                f" #turn = {len(turns) - 1}. (ignored)"
            )

return dict(
    input_ids=input_ids,
    labels=targets,
    attention_mask=input_ids.ne(tokenizer.pad_token_id),
)

`