| """Distinctive per-model vocabulary via several methods, so we can pick the best surface. |
| (1) log-odds-ratio w/ informative Dirichlet prior (Monroe 2008), word 1-2grams, stopwords removed |
| (2) same, bigrams only (phrases) |
| (3) classifier coefficient top terms""" |
| import json |
| from collections import Counter |
| import numpy as np |
| from sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer |
| from sklearn.linear_model import LogisticRegression |
| rows=[json.loads(l) for l in open("data/texts.jsonl")] |
| rows=[r for r in rows if len(r.get("text","") or "")>=1 and r.get("model") in {"gpt-oss","Qwen","GLM-4","Mag","Gemma"}] |
| MODELS=["gpt-oss","Qwen","GLM-4","Mag","Gemma"] |
| NAME={"gpt-oss":"gpt-oss","Qwen":"Qwen","GLM-4":"GLM-4","Mag":"Magistral","Gemma":"Gemma"} |
| texts=[r["text"] for r in rows];model=np.array([r["model"] for r in rows]) |
| def logodds(ngram): |
| cv=CountVectorizer(ngram_range=ngram,min_df=30,max_features=50000,stop_words="english",token_pattern=r"[A-Za-z'][A-Za-z']+") |
| X=cv.fit_transform(texts);vocab=np.array(cv.get_feature_names_out()) |
| tot=np.asarray(X.sum(0)).ravel();a0=tot.sum();alpha=tot.astype(float);N=tot.sum() |
| out={} |
| for m in MODELS: |
| yi=np.asarray(X[model==m].sum(0)).ravel().astype(float) |
| yr=tot-yi |
| ni=yi.sum();nr=yr.sum() |
| d=np.log((yi+alpha)/(ni+a0-yi-alpha))-np.log((yr+alpha)/(nr+a0-yr-alpha)) |
| var=1.0/(yi+alpha)+1.0/(yr+alpha) |
| z=d/np.sqrt(var) |
| out[m]=vocab[np.argsort(z)[::-1][:12]] |
| return out |
| lo1=logodds((1,2));lo2=logodds((2,2)) |
| print("=== (1) log-odds (word 1-2grams) ===") |
| for m in MODELS:print(f" {NAME[m]:10s}: {', '.join(lo1[m])}") |
| print("\n=== (2) log-odds (bigrams / phrases) ===") |
| for m in MODELS:print(f" {NAME[m]:10s}: {', '.join(lo2[m])}") |
|
|
