bi~ ddlZddlmZmZddlmZddlmZddlm Z m Z m Z m Z m Z ddlZddlZddlmZddlZddlmZmZddlmZe deeZd eee fd efd Z d&d eee eeeffd ed e e e ee fd eeeffdZ! d&d eee eeeffd ed e e e ee fd eeeffdZ"de eee eeeffd e eee eeefffdZ# d'dede e$de ed efdZ% d'dede e$de ed efdZ&d eeefd eeeffdZ'd eee fd eee ffdZ(deee e fde$d eee e ffdZ)GddZ*dejVde$d ejVfdZ,dejVde$d ejVfdZ- d(dede$ded e eee fd ef d!Z. d&ded"e$d e eee fd efd#Z/d eee fd efd$Z0d eee fd eee ffd%Z1y))N) defaultdictdeque)Sequence) takewhile)AnyCallableOptionalTypeVarUnion)Dataset DatasetDict)PreTrainedTokenizerBase DatasetTypeexamplereturncgd}|jDchc] }||vs| }}|rC|j}||}t|tr|d}t|tr d|vrd|vryycc}w)aM Check if the example is in a conversational format. Args: example (`dict[str, Any]`): A single data entry of a dataset. The example can have different keys depending on the dataset type. Returns: `bool`: `True` if the data is in a conversational format, `False` otherwise. Examples: ```python >>> example = {"prompt": [{"role": "user", "content": "What color is the sky?"}]} >>> is_conversational(example) True >>> example = {"prompt": "The sky is"} >>> is_conversational(example) False ``` )promptchosenrejected completionmessagesrrolecontentTF)keyspop isinstancelistdict)rsupported_keyskey example_keysmaybe_messages maybe_messages I/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/trl/data_utils.pyis_conversationalr% s0PN#*<<>KCSN5JCKLK   nd +*1-M-.6]3Jy\iOi Ls A-A- tokenizertoolsc gd}|jDchc] }||vs| }}|dhdhddhhdddhhdfvrtd |d|vr|j|d|d }d|vrF|dd d }|dk(rd}d } n|dk(rd }d} ntd||j|d|| d |} d|vrd|vrX|j|d|dz|d } dj dt dt  | D} | t| d} d|vr\d|vrX|j|d|dz|d } dj dt dt  | D} | t| d}d|vr|j|d|dz|d }dj dt dt  |D} |t| d}n6d|vr|j|d|d } d|vr|j|d|d }i}d|vr|d<d|vr |d<d|vr |d<d|vr|d<d|vr|d<d|vr|d|d<|Scc}w)z Apply a chat template to a conversational example along with the schema for a list of functions in `tools`. For more details, see [`maybe_apply_chat_template`]. )rrrrrlabelrrr>rrrrr>r)rrzInvalid keys in the example: F)r'tokenizeruserT assistantz"Invalid role in the last message: )r'continue_final_messager*add_generation_promptc3&K|] \}}| ywN.0x_s r$ z&apply_chat_template..si41aQic|d|dk(SNrr3r6s r$z%apply_chat_template..qtqt|Nc3&K|] \}}| ywr2r3r4s r$r8z&apply_chat_template..sk41aQkr9c|d|dk(Sr;r3r=s r$r>z%apply_chat_template..r?r@c3&K|] \}}| ywr2r3r4s r$r8z&apply_chat_template..sm41aQmr9c|d|dk(Sr;r3r=s r$r>z%apply_chat_template..r?r@textr))rKeyErrorapply_chat_template ValueErrorjoinrziplen)rr&r'rr r!r last_roler/r.r prompt_chosenrprompt_rejectedrprompt_completionroutputs r$rGrGIs3YN#*<<>KCSN5JCKLK   < ( :) 6|nEFFW001DE\a0b7H%b)&1  $( !%* " + %$) !%) "A)MN N.. H #9"7 / 7 w %99!GH$55UU:M WWi95KSQWYfMg+hiiF"3v;=1F  X%8';;!GJ$77uu<OWWk95KSQWYhMi+jkkF&s6{}5H 7 " ) = =!GL$99QV!>! WWm95KSQWYjMk+lmmF*3v;=9J w 22783DE\a2bF   44WZ5HPU`e4fHFW!v7!x7!xW%zw)|'!'*w McLs H?H?c6t|r t|||S|S)a: If the example is in a conversational format, apply a chat template to it. Args: example (`dict[str, list[dict[str, str]]`): Dictionary representing a single data entry of a conversational dataset. Each data entry can have different keys depending on the dataset type. The supported dataset types are: - Language modeling dataset: `"messages"`. - Prompt-only dataset: `"prompt"`. - Prompt-completion dataset: `"prompt"` and `"completion"`. - Preference dataset: `"prompt"`, `"chosen"`, and `"rejected"`. - Preference dataset with implicit prompt: `"chosen"` and `"rejected"`. - Unpaired preference dataset: `"prompt"`, `"completion"`, and `"label"`. For keys `"messages"`, `"prompt"`, `"chosen"`, `"rejected"`, and `"completion"`, the values are lists of messages, where each message is a dictionary with keys `"role"` and `"content"`. tokenizer (`PreTrainedTokenizerBase`): Tokenizer to apply the chat template with. tools (`list[Union[dict, Callable]]` or `None`, *optional*, defaults to `None`): A list of tools (callable functions) that will be accessible to the model. If the template does not support function calling, this argument will have no effect Returns: `dict[str, str]`: Formatted example with the chat template applied. Notes: - This function does not alter the keys, except for Language modeling dataset, where `"messages"` is replaced by `"text"`. - In case of prompt-only data, if the last role is `"user"`, the generation prompt is added to the prompt. Else, if the last role is `"assistant"`, the final message is continued. Example: ```python >>> from transformers import AutoTokenizer >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-128k-instruct") >>> example = { ... "prompt": [{"role": "user", "content": "What color is the sky?"}], ... "completion": [{"role": "assistant", "content": "It is blue."}], ... } >>> apply_chat_template(example, tokenizer) {'prompt': '<|user|>\nWhat color is the sky?<|end|>\n<|assistant|>\n', 'completion': 'It is blue.<|end|>\n<|endoftext|>'} ``` )r%rG)rr&r's r$maybe_apply_chat_templaterRs!j!"7Iu==r@examplescxt|d}|d|dzdg|zdg|zzd}d|vr|d|dz|d<|S)NrrTF)rr)r)rK)rS batch_sizenew_rowss r$ _unpair_rowrWshXh'(Jx(8J+??*$w';;H8%h/(82DD Or@datasetnum_procdescc:|jtdddg||S)a Unpair a preference dataset. Args: dataset (`Dataset` or `DatasetDict`): Preference dataset to unpair. The dataset must have columns `"chosen"`, `"rejected"` and optionally `"prompt"`. num_proc (`int` or `None`, *optional*, defaults to `None`): Number of processes to use for processing the dataset. desc (`str` or `None`, *optional*, defaults to `None`): Meaningful description to be displayed alongside with the progress bar while mapping examples. Returns: `Dataset`: The unpaired preference dataset. Example: ```python >>> from datasets import Dataset >>> dataset_dict = { ... "prompt": ["The sky is", "The sun is"], ... "chosen": [" blue.", "in the sky."], ... "rejected": [" green.", " in the sea."], ... } >>> dataset = Dataset.from_dict(dataset_dict) >>> dataset = unpair_preference_dataset(dataset) >>> dataset Dataset({ features: ['prompt', 'completion', 'label'], num_rows: 4 }) >>> dataset[0] {'prompt': 'The sky is', 'completion': ' blue.', 'label': True} ``` Trr)batchedremove_columnsrYrZ)maprW)rXrYrZs r$unpair_preference_datasetr_s%P ;;{D(JAWbjqu; vvr@ct|tr*|t|jdj}n |j}d|vrd|vrt |||S|S)a Unpair a preference dataset if it is paired. Args: dataset (`Dataset` or `DatasetDict`): Preference dataset to unpair. The dataset must have columns `"chosen"`, `"rejected"` and optionally `"prompt"`. num_proc (`int` or `None`, *optional*, defaults to `None`): Number of processes to use for processing the dataset. desc (`str` or `None`, *optional*, defaults to `None`): Meaningful description to be displayed alongside with the progress bar while mapping examples. Returns: `Dataset` or `DatasetDict`: The unpaired preference dataset if it was paired, otherwise the original dataset. Example: ```python >>> from datasets import Dataset >>> dataset_dict = { ... "prompt": ["The sky is", "The sun is"], ... "chosen": [" blue.", "in the sky."], ... "rejected": [" green.", " in the sea."], ... } >>> dataset = Dataset.from_dict(dataset_dict) >>> dataset = unpair_preference_dataset(dataset) >>> dataset Dataset({ features: ['prompt', 'completion', 'label'], num_rows: 4 }) >>> dataset[0] {'prompt': 'The sky is', 'completion': ' blue.', 'label': True} ``` rrr)rYrZ)rr rr column_namesr_)rXrYrZras r$maybe_unpair_preference_datasetrbs]P';'tGLLN3A67DD ++ <J,$>(8$OOr@c ttt|dt|dD](}|d||d|k7s|d|dz dk(r|dz}n|dd|d|d|d|ddS)z Extracts the shared prompt from a preference data example, where the prompt is implicit within both the chosen and rejected completions. For more details, see [`maybe_extract_prompt`]. rrr< N)rrr)rangeminrK)ridxs r$extract_promptrhLsSWX./WZ5H1IJK 8 S !WZ%8%= =x q)S0q   (#DS)(#CD)J'- r@cd|vsd|vr|Sd|vr*td|di}td|di}|r|s|s|s|St|d|ddS)a Extracts the shared prompt from a preference data example, where the prompt is implicit within both the chosen and rejected completions. If the example already contains a `"prompt"` key, the function returns the example as is. Else, the function identifies the longest common sequence (prefix) of conversation turns between the "chosen" and "rejected" completions and extracts this as the prompt. It then removes this prompt from the respective "chosen" and "rejected" completions. Args: example (`dict[str, list]`): A dictionary representing a single data entry in the preference dataset. It must contain the keys `"chosen"` and `"rejected"`, where each value is either conversational or standard (`str`). Returns: `dict[str, list]`: A dictionary containing: - `"prompt"`: The longest common prefix between the "chosen" and "rejected" completions. - `"chosen"`: The remainder of the "chosen" completion, with the prompt removed. - `"rejected"`: The remainder of the "rejected" completion, with the prompt removed. Examples: ```python >>> example = { ... "chosen": [ ... {"role": "user", "content": "What color is the sky?"}, ... {"role": "assistant", "content": "It is blue."}, ... ], ... "rejected": [ ... {"role": "user", "content": "What color is the sky?"}, ... {"role": "assistant", "content": "It is green."}, ... ], ... } >>> extract_prompt(example) {'prompt': [{'role': 'user', 'content': 'What color is the sky?'}], 'chosen': [{'role': 'assistant', 'content': 'It is blue.'}], 'rejected': [{'role': 'assistant', 'content': 'It is green.'}]} ``` Or, with the `map` method of `datasets.Dataset`: ```python >>> from trl import extract_prompt >>> from datasets import Dataset >>> dataset_dict = { ... "chosen": [ ... [ ... {"role": "user", "content": "What color is the sky?"}, ... {"role": "assistant", "content": "It is blue."}, ... ], ... [ ... {"role": "user", "content": "Where is the sun?"}, ... {"role": "assistant", "content": "In the sky."}, ... ], ... ], ... "rejected": [ ... [ ... {"role": "user", "content": "What color is the sky?"}, ... {"role": "assistant", "content": "It is green."}, ... ], ... [ ... {"role": "user", "content": "Where is the sun?"}, ... {"role": "assistant", "content": "In the sea."}, ... ], ... ], ... } >>> dataset = Dataset.from_dict(dataset_dict) >>> dataset = dataset.map(extract_prompt) >>> dataset[0] {'prompt': [{'role': 'user', 'content': 'What color is the sky?'}], 'chosen': [{'role': 'assistant', 'content': 'It is blue.'}], 'rejected': [{'role': 'assistant', 'content': 'It is green.'}]} ``` rrr)rr)r%rh)r chosen_conv prompt_convs r$maybe_extract_promptrl_sqdw*G";7'783D(EF '783D(EF K[N WX%6GJDWX YYr@ seq_lengthc Rtjdt|jDcic]\}}|t |g}}}|jDcic]0\}}|t dt ||Dcgc] }||||z c}2}}}}|Scc}}wcc}wcc}}}w)a Pack examples into chunks of size `seq_length`. Args: examples (`dict[str, list[list]]`): Dictionary of examples with keys as strings and values as lists of lists. seq_length (`int`): Maximum sequence length. Returns: `dict[str, list[list]]`: Dictionary of examples with keys as strings and values as lists of lists. Example: ```python >>> from trl import pack_examples >>> examples = { ... "input_ids": [[1, 2, 3], [4, 5, 6, 7], [8]], ... "attention_mask": [[0, 1, 1], [0, 0, 1, 1], [1]], ... } >>> pack_examples(examples, seq_length=5) {'input_ids': [[1, 2, 3, 4, 5], [6, 7, 8]], 'attention_mask': [[0, 1, 1, 0, 0], [1, 1, 1]]} >>> pack_examples(examples, seq_length=2) {'input_ids': [[1, 2], [3, 4], [5, 6], [7, 8]], 'attention_mask': [[0, 1], [1, 0], [0, 1], [1, 1]]} ``` zo`pack_examples` is deprecated and will be removed in version 0.20.0. Use `pack_dataset` with a dataset instead.r)warningswarnDeprecationWarningitemssumrerK)rSrmkvis r$ pack_examplesrws: MM  +3..*:;$!Q3q": ;H;]e]k]k]mnnUYUVXY5CFJ3OPaAa!j.)PPnHn O<PnsB!B"=B B"B"c.eZdZdZdefdZdZdZdZy) _SegmentTreea  A segment tree data structure that, when initialized as `_SegmentTree(maxval)`, efficiently finds the next larger value for a given input within the range [1, maxval]. See [Fewer Truncations Improve Language Modeling](https://arxiv.org/abs/2404.10830) for more details. maxvalc.||_dgd|zz|_y)Nrrztree)selfrzs r$__init__z_SegmentTree.__init__s C1v:& r@cd|cxkr|jksJJ|j|zdz }||j|<|dkDrI|dz}|j|dz|j|dzdz}}||k\r|n||j|<|dkDrHyyr;r}rvalrvleftrights r$addz_SegmentTree.adds3%$++%%%%% KK#  ! ! !e !GA))AF+TYYQ!|-D%D#'5=4eDIIaL !er@cd|cxkr|jksJJ|j|zdz }d|j|<|dkDrI|dz}|j|dz|j|dzdz}}||k\r|n||j|<|dkDrHyyr;r}rs r$removez_SegmentTree.removes3%$++%%%%% KK#  ! ! !e !GA))AF+TYYQ!|-D%D#'5=4eDIIaL !er@cd|cxkr|jksJJd}||jkr3|j|dz|k\r|dz}n|dzdz}||jkr3|j|Sr;r})rrrvs r$searchz_SegmentTree.searchsw3%$++%%%%% $++oyya C'F!VqL $++o yy|r@N) __name__ __module__ __qualname____doc__intrrrrr3r@r$ryrys"'s'<<r@ryc |d}|jDcgc]}ttt|!}}t j ||dj }|jd|}g}d}t|jD]\}} tjj| j s)tjj| j rtj| d|} ||}|j!| tj"j%||j&}t)j*t|} |Jtj,tj.||j1| } | j3dd} t5|} | j7|t9t:} g}t=| j?djA| j?d jAD]\}}| jC|}||kr| |jE}ngdd }|j!||d j!||d xx|z cc<||kr| |s| jG|||z }| |j!||dkDs| j7|tjH||Dcgc]}|d D]}|c}}}t)j dg|Dcgc]}|d  c}z}t)jJ|}g}|jD]} t| jLd k(sJ| jLd} tjj| j s)tjj| j rX| jNj jQ}t | j%|jS|| jT} |j!| tj"j%||j&Scc}wcc}}wcc}w) zFPack sequences in a pyarrow Table using First Fit Decreasing strategy. input_ids)type position_idsNrnames descending)byr<)idslengthrr)+ to_pylistrrerKpaarrayr append_column enumeratecolumnspyarrowtypesis_list is_large_listpc list_sliceappendTable from_arraysranparange make_structlist_value_lengthcombine_chunkssortryrrrrJfieldto_numpyrpopleftrtakecumsumchunksoffsetsto_pandas_dtypeastypevalues)rSrmrsequenceposition_ids_pythonposition_ids_arrayrlist_column_idxrgcolumnrlengths segment_tree space_to_binbinsrspacebinid_rdtypes r$ _pack_ffdrs%IFOFYFYF[\(4c(m 45\\"5H[?G&\--gv>Fv  88  x/D/D  EEr@strategy map_kwargsc|i}|jd}|dk(r|jtfdd|id|}n1|dk(r|jtfdd|id|}nt d|d |jd}|S) a Pack sequences in a dataset into chunks of size `seq_length`. Args: dataset (`Dataset` or `DatasetDict`): Dataset to pack seq_length (`int`): Target sequence length to pack to. strategy (`str`, *optional*, defaults to `"ffd"`): Packing strategy to use. Can be either: - `"ffd"` (First Fit Decreasing): Slower but preserves sequence boundaries. Sequences are never cut in the middle. - `"wrapped"`: Faster but more aggressive. Ignores sequence boundaries and will cut sequences in the middle to completely fill each packed sequence with data. map_kwargs (`dict` or `None`, *optional*, defaults to `None`): Additional keyword arguments to pass to the dataset's map method when packing examples. Returns: `Dataset` or `DatasetDict`: The dataset with packed sequences. The number of examples may decrease as sequences are combined. Example: ```python >>> from datasets import Dataset >>> from trl import pack_dataset >>> examples = { ... "input_ids": [[1, 2, 3], [4, 5], [6, 7, 8], [9]], ... "attention_mask": [[1, 1, 0], [1, 0], [1, 0, 0], [1]], ... } >>> dataset = Dataset.from_dict(examples) >>> packed_dataset = pack_dataset(dataset, seq_length=4, strategy="ffd") >>> packed_dataset[:] {'input_ids': [[1, 2, 3, 9], [6, 7, 8, 4, 5]], 'attention_mask': [[1, 1, 0, 1], [1, 0, 0, 1, 0]]} ``` NarrowffdTrm)r\ fn_kwargswrappedzInvalid packing strategy: z. Use 'ffd' or 'wrapped'.) with_formatr^rrrH)rXrmrrs r$ pack_datasetrasR !!'*G5'++ij,PZA[j_ij Y '++mnTlT^E_ncmn5hZ?XYZZ!!$'G Nr@ max_lengthc|i}t|tr>fd}|jd}|j|fddi|}|jd}|Sfd}|j|fddi|}|S)a Truncate sequences in a dataset to a specifed `max_length`. Args: dataset (`Dataset` or `DatasetDict`): Dataset to truncate. seq_length (`int`): Maximum sequence length to truncate to. map_kwargs (`dict` or `None`, *optional*, defaults to `None`): Additional keyword arguments to pass to the dataset's map method when truncating examples. Returns: `Dataset` or `DatasetDict`: The dataset with truncated sequences. Example: ```python >>> from datasets import Dataset >>> examples = { ... "input_ids": [[1, 2, 3], [4, 5, 6, 7], [8]], ... "attention_mask": [[0, 1, 1], [0, 0, 1, 1], [1]], ... } >>> dataset = Dataset.from_dict(examples) >>> truncated_dataset = truncate_dataset(dataset, max_length=2) >>> truncated_dataset[:] {'input_ids': [[1, 2], [4, 5], [8]], 'attention_mask': [[0, 1], [0, 0], [1]]} ``` Nctg}|jD]|}tjj|js)tjj |jrt j|d}|j|~tjj||jS)Nrr) rrrrrrrrrrrrra)rStruncated_columnsrrs r$truncatez"truncate_dataset..truncates " "** 1==((59T9TU[U`U`9a]]61jAF!((0 188''(9AVAV'W Wr@rr\Tci}|jD]1\}}|r%t|dtr|Dcgc]}|d }}|||<3|Scc}w)Nr)rrrr)rStruncated_examplesr rrrs r$rz"truncate_dataset..truncatesf!# '~~/ 1 VjD9:@A3c+:.AFA*0"3' 1& %Bs A )rr rr^)rXrrrs ` r$truncate_datasetrs@ '7# X%%g.'++hCC C%%d+ N &'++     Nr@c|jd}t|tr|d}t|tr d|vrd|vryy)aj Check if the example is in a conversational format (from/value). Note that this format isn't recommended. Prefer the ChatML format (role/content) Args: example (`dict[str, Any]`): A single data entry of a dataset. The example can have different keys depending on the dataset type. Returns: `bool`: `True` if the data is in a conversational Chatformat, `False` otherwise. Examples: ```python >>> example = {"conversations": [{"from": "user", "value": "What color is the sky?"}]} >>> is_conversational_from_value(example) True >>> example = {"conversations": [{"role": "user", "content": "What color is the sky?"}]} >>> is_conversational_from_value(example) False >>> example = {"conversations": "The sky is"}) >>> is_conversational_from_value(example) False ``` conversationsrfromvalueTF)getrrr)rr"r#s r$is_conversational_from_valuersE6[[1N.$'&q) mT *v/F7VcKc r@cdD]i}||vst||ts||}|D]D}t|tsd|vr|jd|d<d|vs1|jd|d<Fkd|vr|jd|d<|S)a3 Convert a conversational dataset with fields `from` and `value` to ChatML format. This function modifies conversational data to align with OpenAI's ChatML format: - Replaces the key `"from"` with `"role"` in message dictionaries. - Replaces the key `"value"` with `"content"` in message dictionaries. - Renames `"conversations"` to `"messages"` for consistency with ChatML. Args: example (`dict[str, list]`): A single data entry containing a list of messages. Returns: `dict[str, list]`: Example reformatted to ChatML style. Example: ```python >>> from trl import maybe_convert_to_chatml >>> example = { ... "conversations": [ ... {"from": "user", "value": "What color is the sky?"}, ... {"from": "assistant", "value": "It is blue."}, ... ] ... } >>> maybe_convert_to_chatml(example) {'messages': [{'role': 'user', 'content': 'What color is the sky?'}, {'role': 'assistant', 'content': 'It is blue.'}]} ``` )rrrrrrrrrrrr)rrrr)rr rmessages r$maybe_convert_to_chatmlrsB[B '>jt<s|H# Bgt,(*1++f*=')-4[[-A *  BB'!%kk/:  Nr@r2)NN)rN)2ro collectionsrrcollections.abcr itertoolsrtypingrrr r r numpyrrrpyarrow.computecomputer pyarrow.typesdatasetsr r transformersrrrstrboolr%rrGrRrWrr_rbrhrlrwryrrrrrrrr3r@r$rs}*$::)0mWk: &tCH~&$&X48] #tDcN++ ,]&] DtX~./ 0] #s(^ ]F488 #tDcN++ ,8&8 DtX~./ 08 #s(^ 8v$tCd38n)=$=>?DcSWX\]`be]eXfSgNgIhDiQU(w (w$,SM(w@H (w(wXQU/ /$,SM/@H //dDh/Dh4G&ZZ$sDy/ZZd39oZZz&Dd4j1&s&tCQUVZQ[OG\&R((V<F<Fc<Fbhh<F~FBHHF#F"((F&jn4 4&)4584NVW[\_ad\dWeNf44pSW> >&)>7?S#X7O>>B#$sCx.#T#L/T#t)_/c4i/r@