File size: 6,453 Bytes
1e80dec | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 | """The flat, long-format merged_models.csv (codebase-audit brief section 8).
One row per model pair x merging method x alignment configuration x benchmark x task x metric,
built by joining summary.csv to merges.csv and models.csv. Nothing here is entered by hand.
Fields no result file records (parent checkpoint hashes, code commit, evaluation timestamp,
dataset revision for most historical runs) are left null, and `limitations` says why.
`result_status` keeps apart the cases that must not be conflated:
evaluated merge and both parents scored under one protocol
parent_baseline_missing merge scored, but a parent was not scored in the same experiment
naive_baseline_missing no naive (unaligned plain-average) twin in the same experiment
naive_baseline_not_applicable the method has no plain-average twin by design (chat vector, OT)
evaluation_failed the source records a failed run
unverified scored, but the source flags the value (e.g. calibration overlap)
Pairs x methods that were never run have no row: "not evaluated" is absence, never a filled value.
"""
import hashlib
import pandas as pd
FLAT = [
# identity and provenance
"experiment_id", "merge_id", "model_family", "model_a", "model_b",
"model_a_checkpoint", "model_b_checkpoint", "model_a_hash", "model_b_hash",
"model_a_parameters", "model_b_parameters", "architecture", "tokenizer_a", "tokenizer_b",
"initialisation_relationship", "training_relationship", "tokenizer_relationship",
"provenance_group",
# method
"merge_method", "alignment_method", "vocabulary_transport", "interpolation_strength",
"method_configuration", "random_seed", "copies_parent", "is_control",
# evaluation
"benchmark", "task", "split", "dataset_revision", "evaluation_samples", "evaluation_metric",
"metric_direction", "model_a_score", "model_b_score", "better_parent_score",
"naive_merge_score", "merged_model_score", "merge_loss", "merge_gain", "merge_success",
"alignment_gain",
# reproducibility
"code_commit", "config_hash", "evaluation_timestamp", "checkpoint_path",
"evaluation_status", "result_status", "limitations", "notes", "source_file",
]
NO_NAIVE_REFERENCE = {"chat_vector", "ot_fusion"}
NOT_RECORDED = ("model_a_hash, model_b_hash, code_commit and evaluation_timestamp are not recorded "
"in the historical result files")
def _status(r):
if str(r["evaluation_status"]) == "failed":
return "evaluation_failed"
if pd.isna(r["best_parent_value"]):
return "parent_baseline_missing"
if str(r["evaluation_status"]) in ("unverified", "partial"):
return "unverified"
if pd.isna(r["naive_value"]):
# chat-vector transfer and OT fusion have no plain-average twin by design
if r["merge_method"] in NO_NAIVE_REFERENCE:
return "naive_baseline_not_applicable"
return "naive_baseline_missing"
return "evaluated"
def flatten(summary: pd.DataFrame, merges: pd.DataFrame, models: pd.DataFrame,
evals: pd.DataFrame) -> pd.DataFrame:
m = merges.set_index("merge_id")
mod = models.set_index("model_id")
# per-measurement metadata from the merge's own evaluation row
ev = evals[evals["merge_id"].fillna("") != ""].drop_duplicates(
["merge_id", "task_name", "dataset_config", "split", "metric_name"]).set_index(
["merge_id", "task_name", "dataset_config", "split", "metric_name"])
def g(df, key, col):
try:
v = df.at[key, col]
except KeyError:
return None
return None if (isinstance(v, float) and pd.isna(v)) else v
rows = []
for _, s in summary.iterrows():
r = m.loc[s["merge_id"]]
a, b = r["parent_a"], r["parent_b"]
ek = (s["merge_id"], s["task_name"], s["dataset_config"], s["split"], s["metric_name"])
cfg = "" if pd.isna(r["merge_config"]) else str(r["merge_config"])
out = {
"experiment_id": s["experiment_id"], "merge_id": s["merge_id"],
"model_family": r["experiment_family"], "model_a": a, "model_b": b,
"model_a_checkpoint": g(mod, a, "checkpoint_uri"), "model_b_checkpoint": g(mod, b, "checkpoint_uri"),
"model_a_hash": None, "model_b_hash": None,
"model_a_parameters": g(mod, a, "parameters"), "model_b_parameters": g(mod, b, "parameters"),
"architecture": r["model_architecture"],
"tokenizer_a": g(mod, a, "tokenizer"), "tokenizer_b": g(mod, b, "tokenizer"),
"initialisation_relationship": r["initialisation_relationship"],
"training_relationship": r["training_relationship"],
"tokenizer_relationship": r["tokenizer_relationship"],
"provenance_group": f"{r['initialisation_relationship']}|{r['training_relationship']}|{r['tokenizer_relationship']}",
"merge_method": r["merge_method"], "alignment_method": r["alignment_method"],
"vocabulary_transport": r["vocabulary_transport"], "interpolation_strength": r["merge_alpha"],
"method_configuration": cfg, "random_seed": r["experiment_seed"],
"copies_parent": r["copies_parent"], "is_control": r["is_control"],
"benchmark": s["task_name"], "task": s["dataset_config"], "split": s["split"],
"dataset_revision": None, "evaluation_samples": g(ev, ek, "n_examples"),
"evaluation_metric": s["metric_name"], "metric_direction": s["metric_direction"],
"model_a_score": s["parent_a_value"], "model_b_score": s["parent_b_value"],
"better_parent_score": s["best_parent_value"], "naive_merge_score": s["naive_value"],
"merged_model_score": s["method_value"], "merge_loss": s["method_merge_loss"],
"merge_gain": s["merge_gain"], "merge_success": s["merge_success"],
"alignment_gain": s["alignment_gain"],
"code_commit": None, "config_hash": hashlib.sha1(cfg.encode()).hexdigest()[:12] if cfg else None,
"evaluation_timestamp": None, "checkpoint_path": r["checkpoint_uri"],
"evaluation_status": g(ev, ek, "evaluation_status") or r["evaluation_status"],
"limitations": NOT_RECORDED, "notes": r["notes"], "source_file": r["source_file"],
}
out["result_status"] = _status(s)
rows.append(out)
return pd.DataFrame(rows, columns=FLAT)
|