suchirsalhan's picture
Initial release: summer merge experiments (18,171 merges, 64,180 evaluations)
1e80dec verified
Raw History Blame Contribute Delete
6.16 kB
"""Schema for the model-merging dataset: three linked tables plus a derived summary.
models.csv one row per unique checkpoint (parent, merged, control)
merges.csv one row per merge experiment (pair x method x alignment x config)
evaluations.csv long format: one row per model x task x config x metric, in one experiment
Comparisons are made only *within* an experiment_id. A parent's score is joined to a merge's
score when the two rows share experiment_id, task_name, dataset_config, split and metric_name,
so a parent evaluated under a different protocol is never used as a merge's baseline.
Every row records the result file it came from (source_file). Values that no file records are
left empty, never inferred.
"""
MODELS = [
"model_id", # HF repo id when one exists, else "local:<family>/<name>"
"model_role", # parent | merged | control | reference (ensemble, oracle)
"experiment_family",
"model_architecture", # pico | gpt_neox | llama | gemma | ... | unknown
"parameters", # integer count, or empty
"tokenizer",
"initialisation", # seed / init id as recorded, or empty
"training_data", # corpus / language / curriculum as recorded
"training_steps",
"checkpoint_uri", # hf://<id>[@rev] | path | empty
"is_published", # True if the checkpoint exists on the Hub
"source_file",
]
MERGES = [
"merge_id", # stable, see make_merge_id
"experiment_id", # the evaluation context comparisons are made in
"experiment_family", # beetle_published | beetle_controlled | polypythia | toksuite |
# cross_architecture | llama_forks | mergebench | humanscale
"parent_a",
"parent_b",
"merged_model_id", # checkpoint id if materialised/published, else empty
"merge_method", # canonical names: METHODS
"alignment_method", # canonical names: ALIGNMENTS
"vocabulary_transport", # True | False
"model_architecture",
"parameters",
"initialisation_relationship", # shared | independent | unknown
"training_relationship", # shared_history | divergent_history | independent | unknown
"tokenizer_relationship", # shared | different | unknown
"merge_alpha",
"merge_config", # JSON string with every recorded hyperparameter
"checkpoint_uri",
"experiment_seed",
"is_control", # True for shuffled/permuted/random controls
"copies_parent", # True when the operator output is (effectively) one parent's tensors
"evaluation_status", # complete | partial | failed | unverified
"notes",
"source_file",
]
EVALUATIONS = [
"experiment_id",
"merge_id", # empty for parent / reference rows
"model_id",
"model_role", # parent_a | parent_b | merged | control | reference
"merge_method", # empty for parents
"alignment_method",
"task_name", # canonical: TASKS keys
"task_category", # TASKS values
"eval_group", # core | family_specific | exploratory
"dataset_name",
"dataset_config", # subset / language / eval_langs as recorded
"split",
"metric_name", # accuracy | nll | perplexity | exact_match | strict_acc | loose_acc ...
"metric_value",
"metric_direction", # higher | lower
"n_examples",
"evaluation_seed",
"evaluation_status", # verified | provisional | failed
"evaluation_notes",
"source_file",
]
SUMMARY = [
"merge_id", "experiment_id", "experiment_family", "merge_method", "alignment_method",
"vocabulary_transport", "initialisation_relationship", "tokenizer_relationship", "parameters",
"task_name", "dataset_config", "split", "metric_name", "metric_direction",
"parent_a_value", "parent_b_value", "best_parent_value", "naive_value", "method_value",
"naive_merge_loss", # NLL: L_naive - min(L_A, L_B); accuracy: max(A, B) - acc_naive
"method_merge_loss", # same, for this method (positive = worse than best parent)
"merge_gain", # naive_merge_loss - method_merge_loss (positive = method helped)
"merge_success", # method beats the best parent (strictly)
"beats_parent_a", "beats_parent_b",
"task_retention", # method_value / best_parent_value (accuracy only)
"alignment_gain", # unaligned-twin loss - aligned loss, same method and pair
"copies_parent", "is_control", "evaluation_status",
]
METHODS = { # canonical merge_method values
"average", "task_arithmetic", "ties", "dare", "qmd", "slerp", "csp", "svc", "opcm",
"ot_fusion", "chat_vector", "interpolation", "ensemble", "other",
}
ALIGNMENTS = {"none", "permutation", "ot", "procrustes", "vocab_transport", "other"}
TASKS = { # task_name -> task_category
"blimp": "linguistic", "multiblimp": "multilingual", "xblimp": "multilingual",
"wikitext": "language_modelling", "fineweb": "language_modelling", "lm_nll": "language_modelling",
"flores": "multilingual", "ifeval": "instruction_following", "pythia_suite": "general",
"arc_easy": "general", "arc_challenge": "general", "piqa": "general", "hellaswag": "general",
"lambada": "general", "sciq": "general", "winogrande": "general",
"gsm8k": "reasoning", "math": "reasoning", "mmlu": "general", "other": "other",
}
CORE_TASKS = {"blimp", "wikitext", "lm_nll", "pythia_suite", "arc_easy", "piqa", "lambada", "sciq"}
LOWER_IS_BETTER = {"nll", "perplexity", "loss", "bits_per_byte"}
def make_merge_id(family, parent_a, parent_b, method, alignment, config=""):
"""Stable id: two merges differing in any config value are different experiments."""
import hashlib
key = "|".join(map(str, (family, parent_a, parent_b, method, alignment, config)))
return f"{family}:{hashlib.sha1(key.encode()).hexdigest()[:12]}"
def direction(metric_name):
return "lower" if metric_name in LOWER_IS_BETTER else "higher"