| """Model style directions (centroids), their pairwise comparison + stability, |
| and their cosine distance to an 'engagingness' direction (linear engagement probe). |
| Style space = agent-level TF-IDF -> SVD (K=30), axis 1 (language) dropped.""" |
| import sqlite3, json |
| from collections import defaultdict |
| from pathlib import Path |
| import numpy as np |
| from numpy.linalg import lstsq, norm |
| from sklearn.feature_extraction.text import TfidfVectorizer |
| from sklearn.decomposition import TruncatedSVD |
| RUNS=Path("runs") |
| OUT=Path("out/style_geometry.md") |
| SHORT={"Qwen/Qwen3-32B":"Qwen","openai/gpt-oss-20b":"gpt-oss","mistralai/Magistral-Small-2509":"Mag","zai-org/GLM-4-32B-0414":"GLM-4","google/gemma-4-31B-it":"Gemma"} |
| K=30; MODELS=["gpt-oss","Qwen","GLM-4","Mag","Gemma"] |
| ag=[] |
| for db in sorted(RUNS.glob("sweep_*.db")): |
| if any(x in db.name for x in ["archive","olm","hermes","glm47","pilot_","snapshot","qwq"]):continue |
| mp=str(db).replace(".db",".metadata.json") |
| if not Path(mp).exists():continue |
| try:amm=json.load(open(mp)).get("agent_model_map",{}) |
| except:continue |
| if not amm:continue |
| fam={int(k)+1:SHORT.get(v["model"]) for k,v in amm.items()} |
| if any(v is None for v in fam.values()):continue |
| try: |
| c=sqlite3.connect(db);posts=c.execute("SELECT post_id,user_id,content FROM post").fetchall();coms=c.execute("SELECT post_id,user_id FROM comment").fetchall();c.close() |
| except:continue |
| if len(coms)<200:continue |
| owner={p:u for p,u,_ in posts};ind=defaultdict(int) |
| for p,cu in coms: |
| o=owner.get(p) |
| if o in fam and o!=cu:ind[o]+=1 |
| txt=defaultdict(list) |
| for p,u,cont in posts: |
| if u in fam and cont:txt[u].append(cont) |
| for u,tl in txt.items(): |
| if tl:ag.append({"run":db.stem,"model":fam[u],"indeg":ind.get(u,0),"text":" ".join(tl)}) |
| n=len(ag);print(n,"agents") |
| vc=TfidfVectorizer(ngram_range=(1,2),analyzer="word",min_df=10,max_features=20000,sublinear_tf=True,norm="l2",stop_words="english",token_pattern=r"[A-Za-z'][A-Za-z']+") |
| X=vc.fit_transform([a["text"]for a in ag]) |
| S=TruncatedSVD(K,random_state=0).fit_transform(X) |
| Z=((S-S.mean(0))/(S.std(0)+1e-9))[:,1:] |
| model=np.array([a["model"]for a in ag]);runs=np.array([a["run"]for a in ag]) |
| y=np.zeros(n);bd=defaultdict(list) |
| for i,a in enumerate(ag):bd[a["run"]].append(i) |
| for r,idx in bd.items(): |
| v=np.array([ag[i]["indeg"]for i in idx],float);z=(v-v.mean())/(v.std()+1e-9) |
| for k,i in enumerate(idx):y[i]=z[k] |
| def u(x):return x/(norm(x)+1e-12) |
| cos=lambda a,b:float(u(a)@u(b)) |
| cent={m:Z[model==m].mean(0) for m in MODELS} |
| |
| A=np.column_stack([np.ones(n),Z]);beta=lstsq(A,y,rcond=None)[0][1:];e=u(beta) |
| o=["# Style geometry: model directions, stability, and the engagingness vector\n", |
| "Style space = agent TF-IDF -> SVD(30), axis 1 (language) dropped (29 dims).\n", |
| "## 1. Model directions (centroids) -- pairwise cosine", |
| "| | "+" | ".join(MODELS)+" |","|"+"---|"*(len(MODELS)+1)] |
| for m in MODELS:o.append(f"| {m} | "+" | ".join(f"{cos(cent[m],cent[m2]):+.2f}" for m2 in MODELS)+" |") |
| |
| rng=np.random.default_rng(0);uruns=np.array(sorted(set(runs)));B=60 |
| selfc={m:[] for m in MODELS};e_self=[] |
| for _ in range(B): |
| rng.shuffle(uruns);Aset=set(uruns[:len(uruns)//2]);am=np.array([r in Aset for r in runs]) |
| for m in MODELS: |
| a1=(model==m)&am;b1=(model==m)&(~am) |
| if a1.sum()<10 or b1.sum()<10:continue |
| selfc[m].append(cos(Z[a1].mean(0),Z[b1].mean(0))) |
| |
| bA=lstsq(np.column_stack([np.ones(am.sum()),Z[am]]),y[am],rcond=None)[0][1:] |
| bB=lstsq(np.column_stack([np.ones((~am).sum()),Z[~am]]),y[~am],rcond=None)[0][1:] |
| e_self.append(cos(bA,bB)) |
| o.append("\n## 2. Direction stability (cosine between centroids from disjoint run-halves, mean over 60 splits)") |
| o.append("| model | self-cosine (1=perfectly stable) |");o.append("|---|---|") |
| for m in MODELS:o.append(f"| {m} | {np.mean(selfc[m]):.2f} |") |
| o.append(f"\nEngagingness vector self-cosine across run-halves: **{np.mean(e_self):.2f}** " |
| f"(stability of the engagement direction itself).") |
| |
| o.append("\n## 3. Cosine of each model direction to the engagingness vector (higher = style closer to 'what engages')") |
| o.append("| model | cos(centroid, engagingness) |");o.append("|---|---|") |
| ranked=sorted(MODELS,key=lambda m:cos(cent[m],e),reverse=True) |
| for m in ranked:o.append(f"| {m} | {cos(cent[m],e):+.2f} |") |
| o.append(f"\nClosest to the engagingness direction: **{ranked[0]}**. " |
| "If that is gpt-oss, the engaging style direction coincides with the attractor's style.") |
| OUT.write_text("\n".join(o)+"\n");print("\n".join(o));print("wrote",OUT) |
|
|
