lzhang472/step-probe-hidden-states

Step-level hidden states for reasoning-faithfulness probes

Hidden states of Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct and Qwen3-8B, extracted at every reasoning step of three released step-annotated chain-of-thought datasets -- FaithCoT-Bench, ProcessBench and GRACE -- plus token-level paths inside each step for FaithCoT-Bench and GRACE. They were built to ask whether a model's hidden state carries step-level label information beyond cheap observable baselines (step position, wording, token confidence), and whether path signatures of hidden-state trajectories add anything beyond the plain state.

Produced by the High-Order Structural LLM Reasoning project (KCL / UCL), September 2026. About 13 GB, float16 NumPy arrays plus Parquet / JSON metadata.

What is in it

folder	dataset	generator (= model whose states these are)	policy	rows (steps)	array shape per file
faithcot_hidden/	FaithCoT-Bench	Qwen2.5-7B-Instruct	on-policy	2362 (341 traces)	[2362, 29, 3584]
faithcot_hidden_llama31/	FaithCoT-Bench	Llama-3.1-8B-Instruct	on-policy	2624 (340 traces)	[2624, 33, 4096]
processbench_hidden/	ProcessBench	Qwen2.5-7B-Instruct / Llama-3.1-8B-Instruct	approximately on-policy	2801 / 4071	[n, 29, 3584] / [n, 33, 4096]
grace_hidden/	GRACE (train + test)	Qwen2.5-7B-Instruct / Llama-3.1-8B-Instruct / Qwen3-8B	approximately on-policy	2049 / 4132 / 1991	[n, 29, 3584] / [n, 33, 4096] / [n, 37, 4096]
faithcot_tok_{qwen25,llama31}/	FaithCoT-Bench, token level	as above	on-policy	125004 / 164142 tokens	[tokens, 4, 512]
faithcot_tok_{qwen25,llama31}_shallow/	same, shallow layers	as above	on-policy	same	[tokens, 4, 512] / [tokens, 3, 512]
grace_tok/	GRACE, token level	Qwen2.5-7B-Instruct / Llama-3.1-8B-Instruct	approximately on-policy	117626 / 260560 tokens	[tokens, 2, 512]

On-policy means the released text was reproduced token by token under the recovered generation prompt (greedy agreement 0.984-0.996), so the states are those of the generation. Approximately on-policy means the text was written by the same model but sampled; the prompt was chosen by response likelihood among candidates and cannot be verified, and ProcessBench's release re-segmented the text.

Step-level files (*_hidden/)

For each {dataset}_{tag} (tags: qwen25, llama31, qwen3_8b):

file	content
{dataset}_{tag}_mean.npy	[steps, layers + 1, hidden], float16: mean of the residual-stream states over the step's tokens, at the embedding output (index 0) and after every decoder layer
{dataset}_{tag}_last.npy	same shape: the state at the step's last token
{dataset}_{tag}_uncertainty.parquet	per step: mean_logp, min_logp (token log-probabilities of the released tokens), mean_ent, max_ent (predictive entropy)
{dataset}_{tag}_meta.json	model, layer count, hidden size, dtype, prompt format and its ranking (approximately on-policy sets)
{dataset}_{tag}_index.parquet	one row per array row: trace_id, step_index, n_steps, token span (tok_start, tok_end), ok, has_step_annotation, step_unfaithful (step label; empty where the step is unlabelled), unfaithfulness (trace label); step text: see below

Row i of the arrays is row i of the index. Keys are trace_id + step_index (1-based), matching the released datasets: FaithCoT-Bench {source}/{generator directory}/{question id}, ProcessBench {split}/{id}, GRACE {train|test}/{id}.

Token-level files (faithcot_tok_*, grace_tok/)

file	content
*_states.npy	[tokens, n_layers, 512], float16: every generated token's state at the layers listed in *_meta.json, after a fixed Gaussian random projection to 512 dimensions (seed 0)
*_projection.pt	the projection matrix (FaithCoT-Bench folders)
*_steps.parquet	trace_id, step_index, tok_start, tok_end into the token array
*_traces.parquet	trace_id, offset and token count per trace (FaithCoT-Bench also dataset, n_steps, the trace-level label)
faithcot_tokcurve_*.parquet	early-answering curve at every token of steps where the answer moves: q (probability of the final answer when the chain is cut at that token), p_top, and the released step-end values

Layers: Qwen2.5 [8, 16, 20, 24] and shallow [0, 2, 4, 8]; Llama [9, 18, 23, 27] and shallow [0, 2, 4]; GRACE [16, 20] for both models.

Index files and step text

• ProcessBench indexes include step_text (the dataset is Apache-2.0).
• FaithCoT-Bench and GRACE indexes are text-free: FaithCoT-Bench's repository has no license and GRACE's license could not be confirmed, so their text is not redistributed. Each row keeps step_text_sha1 and step_text_chars instead. The text is exactly recoverable from the original releases -- FaithCoT-Bench: sample_0["step_<i>"] of <source>/<generator>/response_<qid>.json; GRACE: steps[i - 1]["text"] of record <split>/<id> -- and the project script src/data/restore_step_text.py does this and checks every row against the hash (verified: all 13,158 rows round-trip exactly).

Loading

import numpy as np, pandas as pd
from huggingface_hub import snapshot_download

root = snapshot_download("lzhang472/step-probe-hidden-states", repo_type="dataset",
                         allow_patterns=["grace_hidden/grace_qwen25_*"])
H = np.load(f"{root}/grace_hidden/grace_qwen25_mean.npy", mmap_mode="r")   # [2049, 29, 3584]
state_L20 = np.asarray(H[:, 20, :], dtype=np.float32)

Sources and licenses

component	source	license
FaithCoT-Bench	Shen et al., ICLR 2026, arXiv:2510.04040, github.com/se7esx/FaithCoT-BENCH	no license file in the repository
ProcessBench	Zheng et al., 2024, arXiv:2412.06559, Qwen/ProcessBench	Apache-2.0 (step text included here)
GRACE	arXiv:2606.16151	not confirmed
Qwen2.5-7B-Instruct, Qwen3-8B	Qwen team	Apache-2.0
Llama-3.1-8B-Instruct	Meta	Llama 3.1 Community License. Built with Llama.

The arrays here are derived model activations, not text. If you use them, cite the source datasets and models above.