File size: 1,322 Bytes
73f7e6f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Collect eval_synth.py outputs (out/evals/*.json) into out/stats/evals.json for charts and the card."""
import glob
import json
import os

OUT = os.environ.get("OPUSDB_OUT", "out")
META = {  # file stem -> (label, params in billions, kind, base model, train examples, note)
    "smollm2-135m": ("SmolLM2-135M-Instruct", 0.135, "baseline", None, 0, ""),
    "SmolLM2-360M-Instruct": ("SmolLM2-360M-Instruct", 0.36, "baseline", None, 0, ""),
    "Qwen2.5-0.5B-Instruct": ("Qwen2.5-0.5B-Instruct", 0.5, "baseline", None, 0, ""),
    "Qwen2.5-1.5B-Instruct": ("Qwen2.5-1.5B-Instruct", 1.5, "baseline", None, 0, ""),
    "smollm2-135m-opus": ("SmolLM2-135M-Instruct + OPUS-SFT", 0.135, "opus-sft", "HuggingFaceTB/SmolLM2-135M-Instruct",
                          4000, "1 epoch on `opus-synth` rows, CPU, lr 1e-4"),
}

rows = []
for f in sorted(glob.glob(os.path.join(OUT, "evals", "*.json"))):
    stem = os.path.basename(f)[:-5]
    if stem not in META:
        continue
    s = json.load(open(f))["summary"]
    label, p, kind, base, n_train, note = META[stem]
    rows.append({**s, "label": label, "params_b": p, "kind": kind, "base_model": base, "train_examples": n_train,
                 "train_note": note})
json.dump(rows, open(os.path.join(OUT, "stats", "evals.json"), "w"), indent=2)
print(len(rows), "eval rows")