tom-yx-wang/verb-overgen-temp

Shared Dataset: overgen_results-260413T2011.parquet

Columns

Column	Type	Description
run_name	string	Each run is a unique training, 100 in total.
run_seed	int32	One of [189, 352, 559, 597, 706]. Concerns data shuffling, sampling, model initialization.
run_subset	string	One of ["remove_preemption-all_verb", "remove_preemption-per_verb", "remove_entrenchment-all_verb", "remove_entrenchment-per_verb"]. Remove preemption may be abbr as -P, remove entrenchment (entrenchment minus preemption) as +P.
epoch	int32	One of [0, 1, 2]. Epoch of the training when the evaluation is done.
verb	string	The verb in the eval sentence. One of ["cry", "laugh", "smile", "sing", "sleep", "go", "sneeze", "blink", "look"]. In per-verb runs, this is guaranteed to be the same as the verb manipulated during training.
index_sentence	int64	NOT CONSECUTIVE. If this index is different, the model receives a different prompt. e.g. "Tom made Jerry laugh", "Tom laughed Jerry", and "Tom told Jerry a story. Tom laughed Jerry." have DIFFERENT index_sentence.
index_verb_context	int64	NOT CONSECUTIVE. If this index is different, the verb appears in a different context (e.g. different subjects, different context sentences). e.g. "Tom made Jerry laugh", "Tom laughed Jerry", and "Tom told Jerry a story. Tom laughed Jerry." have SAME index_verb_context. The index is verb-specific: the same index across different verbs DOES NOT imply the same context.
sentence_construction	string	One of ["causative", "transitive", "context_causative", "context_transitive"]. The kind of sentence construction.
sentence	string	Raw sentence, not lemmatized. Conversion to past tense not implemented yet.
lemma	list[string]	Lemmatized sentence provided by spacy.
input_ids	list[int]	Token IDs from the word-level tokenizer, tokenized on the lemmas. 1 = BOS_TOKEN; 0, 2, 3 not used; 4 = period. All sentences start with BOS. Context and transitive/causative sentences are concatenated with BOS as separator.
token_prob	list[float32]	Next-token prediction probabilities for each token in input_ids. token_prob[0] is the probability on the first token after BOS. 1 element shorter than input_ids.
nll	float32	Negative log likelihood of the final sentence. For context sentences, only the latter sentence is scored (e.g. <s> Tom told Jerry a story . <s> Tom made Jerry laugh . scores Tom made Jerry laugh .).