nla-thought-anchors-rollouts

Sentence-level NLA rollout dataset for Qwen2.5-7B-Instruct on GSM8K test set.

For each activation vector (layer 20, 1319 examples), 40 NLA descriptions were sampled from the actor kitft/nla-qwen2.5-7b-L20-av (temperature=0.7). Each description was split into sentences and scored with the NLA critic kitft/nla-qwen2.5-7b-L20-ar.

Part of the nla-thought-anchors project.

Files

File	Extraction point	Rows
hash.parquet	#### token (model finished CoT, about to write answer)	175,240
answer.parquet	Answer digit token (answer now in context)	~175,000

Schema

Column	Type	Description
example_idx	int32	Index in GSM8K test split
is_correct	bool	Whether model answer matched gold
gold_answer	string	Ground-truth answer
rollout_idx	int32	0–39 — which of the 40 sampled descriptions
sentence_idx	int32	Sentence position within this rollout
sentence_text	string	The sentence text
sentence_embedding	list<float32> len=384	MiniLM-L6-v2 embedding
cos_sim_to_full_nla	float32	Cos-sim of sentence to its own rollout's full description
critic_cos_sim	float32	NLA critic reconstruction cos-sim vs original activation
critic_mse	float32	NLA critic reconstruction MSE vs original activation
contains_gold_answer	bool	Sentence contains the gold answer number
contains_pred_answer	bool	Sentence contains the predicted answer number

Key findings

Reconstruction quality by correctness (hash position):

Group	Mean critic cos-sim
Correct examples	0.604
Incorrect examples	0.416

Incorrect examples reconstruct ~35% worse — the activation at #### is qualitatively different when the model gets the answer wrong.

Thought anchor analysis: sentences that are stable across all 40 rollouts (high cos-sim to cross-rollout centroid) reconstruct significantly better than peripheral sentences:

Group	Correct	Incorrect
Peripheral sentences (Q1 anchor score)	0.503	0.366
Anchor sentences (Q4 anchor score)	0.661	0.673

The most striking result: anchor sentences from incorrect examples (0.673) reconstruct better than peripheral sentences from correct examples (0.503) — the stable recurring content is what carries the activation signal regardless of correctness.

Anchor score is not stored in this dataset — compute it post-hoc from sentence_embedding by grouping rows by example_idx, computing the centroid of all embeddings, and taking cos-sim of each sentence to the centroid.

Usage

from datasets import load_dataset
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

ds = load_dataset("Realmbird/nla-thought-anchors-rollouts", split="hash")

# compute anchor scores for each example
import pandas as pd
df = ds.to_pandas()
embs = np.stack(df["sentence_embedding"].values)

anchor_scores = np.zeros(len(df))
for ex_idx, group in df.groupby("example_idx"):
    idx = group.index
    centroid = embs[idx].mean(axis=0)
    centroid /= np.linalg.norm(centroid) + 1e-12
    sims = cosine_similarity(embs[idx], centroid.reshape(1, -1)).flatten()
    anchor_scores[idx] = sims

df["anchor_score"] = anchor_scores

Related

• Step 1 activations (#### token): Realmbird/nla-thought-anchors-hash
• Step 1 activations (answer token): Realmbird/nla-thought-anchors-answer
• Step 2 NLA descriptions: Realmbird/nla-thought-anchors-hash-step2
• NLA actor: kitft/nla-qwen2.5-7b-L20-av
• NLA critic: kitft/nla-qwen2.5-7b-L20-ar