mangesh-ux/thinking-completion-probes

Thinking Completion Probes

An anonymized numerical derivative release for an exploratory study of Qwen3-1.7B: does its residual state at generated token 512 improve a forecast of natural thinking completion within 4096 generated tokens, beyond specified confidence and lexical controls? Completion is not answer correctness.

Code and study · GitHub v1.0.0 release · Methodology · Release/privacy policy

Use this bundle to reproduce the published probe/control comparison or explore alternative analyses on TRAIN. Preserve problem-level grouping and the supplied folds; validation is already published evidence, not a fresh holdout for tuning new claims. Confirmatory extensions need a separately specified, new held-out study. This publication does not launch one.

Scope and original outcome

The original V2-A1 experiment asked about wrong→correct answer recovery. Collection completed: 768 base trajectories, 768 terminal readouts, 3840 checkpoint/readout slots and 3793 available activation matrices, with zero interrupted/failed collection operations. Unavailable slots were preserved. Answer readout/reference evaluation could not establish eligible wrong→correct and wrong→wrong populations. That endpoint is NON_EVALUABLE, not a negative activation result. No original recovery probes or test evaluation were completed.

The separately locked completion follow-up reused existing trajectories, prefixes, confidence features and activations without new GPU inference. It is not evidence of causal reasoning understanding, remaining-token estimation, accuracy-preserving early stopping or inference-cost savings.

Dataset construction and labels

An answer-independent, fixed 160-problem difficulty-7/8 cohort was grouped into 96 TRAIN, 32 validation and 32 unused test problems. Six seeded stochastic trajectories per TRAIN/validation problem yield the released 768 rows.

Split	Rows	Independent problems	Natural-close rows	Full-cap no-close rows	Mixed-outcome problems
TRAIN	576	96	151	425	46
Validation	192	32	67	125	17

The target is conditional on an available, still-open 512-token reasoning prefix. labels=1 means the first natural </think> starts at or after that checkpoint and within the 4096-token generation; labels=0 means a genuine full-cap generation without that close. Already-closed, short, failed, early-EOS or unknown-termination rows are exclusions, not negatives. Historical TRAIN and validation had no primary-checkpoint attrition. The original test split is absent.

Predictors exclude eventual answers, correctness, close position, termination reason and remaining-token count. Features were reconstructed after collection from prefix information: prefix-information-only, not prospectively timestamped pre-outcome online measurements. Labels use the subsequent natural trajectory.

Files and schema

This is a NumPy NPZ/JSON bundle, not a row-text dataset. Use the pinned Python loader/example below; automatic datasets.load_dataset support is not provided. train.npz and validation.npz contain the same fields, with N=576 and 192:

Field	Shape / type	Meaning
activations	[N,28,2048] float32	All zero-based transformer blocks' post-block final-prefix-token residuals, before final normalization
numeric	[N,11] float64	Four structural and seven frozen readout confidence/agreement features
numeric_feature_names	[11] Unicode	Exact numeric column names
lexical	[N,512] float64	Fixed SHA256-bin counts of the last 128 generated prefix tokens; collisions, no order
labels	[N] int8	Natural thinking close (1) versus full-cap no-close (0)
groups, row_ids	[N] Unicode	Anonymized problem grouping and row identifiers
fold_ids	[N] int8	Frozen TRAIN folds 0–4; validation uses −1
topics, difficulty	[N] Unicode / float64	Categorical topic and difficulty 7 or 8

Activations replay the exact prompt plus saved generated-512 prefix without an answer cue. Original BF16 values are represented losslessly as float32. Confidence features come from separate greedy, 64-token counterfactual readouts; their original measurement limitations are retained. The three answer-span fields (mean_answer_token_logprob, min_answer_token_logprob, answer_token_count) are NaN in 566/576 TRAIN rows and all 192 validation rows. Entropy, margin, agreement and availability fields are not thereby all missing. previous_readout_available means both current and preceding boxed strings exist, not that both receipt files exist. Problem-token count excludes the chat template. Missingness is explicitly handled using TRAIN-fitted preprocessing.

fitted-models.json contains frozen model/preprocessor states; heldout-predictions.npz contains paired validation predictions. manifest.json records schemas, support, privacy exclusions and original provenance hashes separately from the newly serialized/anonymized file hashes. Scientific values are preserved; byte equality to the original private NPZ archives is not asserted.

Frozen analysis and result

Models are L2 logistic regressions. A uses structural/topic features; B adds seven readout features; Blex adds hashed prefix counts; C adds one activation layer to B; Clex adds that layer to Blex. Clex versus Blex is primary. Five problem-grouped, difficulty-stratified TRAIN folds select among 128 fixed configurations (640 CV fits), followed by five final TRAIN fits. All preprocessing is fit within each TRAIN partition. The Clex-selected layer is zero-based 15 (block 16); all selected regularization values are 0.001. C uses that layer, selecting only its own regularization. Preserve supplied folds: rehashing anonymous IDs would change the analysis. Validation does not select layers or fit models.

On 32 held-out validation problems, problem-averaged log loss was 0.6687 nats for Blex versus 0.6134 for Clex: baseline-minus-added improvement 0.0553 nats per problem, with a paired problem-cluster 95% interval [0.0212, 0.0888]. The 2000 resamples preserve repeated sampled problems and condition on frozen trained models; they do not include training, layer-selection or exploratory endpoint-selection uncertainty. Secondary comparisons are descriptive and not multiplicity-corrected. The 192 validation rows are not 192 independent problems.

This supports incremental predictive utility relative to these controls, not information absent from the tokens. One model/BF16 backend, one checkpoint, a capped horizon, a weak lexical control and only 32 validation problems limit generalization. Original test remaining unused does not make the post-collection endpoint confirmatory. No causal intervention or deployment policy was tested.

Integrity and provenance

Dataset revision: v1.0.0. The example resolves the tag to an immutable Hub commit and verifies the independently pinned manifest plus every full file hash and byte count before numerical verification.

File	Bytes	SHA256
train.npz	62236053	6490071a17cd7f5779de97645573db885ce55a85191758957a36c7cc782e6fc2
validation.npz	20746215	0a25de9d40bd0171122c98315e45167e3984e8c64d7aede4288d6f46d39c3f89
fitted-models.json	546069	da1b0e502732a2c779cba6b263e0df1094b744aaac5b93f454960d087b5d7843
heldout-predictions.npz	8610	e4190d59a41a4caa437044042d2628f93e79e6109d2a31324ef3869eee476507
manifest.json	5690	fed6a224ef7b6576de47f27cc2d5ef42e171999df57ffc1aa66dfb2e3f431c7f

Historical authority is reasoning-recovery-probes at 27ecb795ea8194a5fe3c3d603a6b775e7f705088. Released code v1.0.0 is pinned to d71422e0f59fddcbff8960c876201301b8be4b4c. The separate GitHub ZIP has SHA256 baea58d3a5156bc70bdf46f0c105cc00ba1bf92db0145051969475215d1d2383; it is not duplicated here. Hashes establish integrity, not scientific validity.

CPU reproduction

Use Python 3.12 (historical stack: 3.12.3), the code's pinned CPU requirements and the separately pinned delivery client. Linux/macOS example; on Windows use .venv\Scripts\Activate.ps1 for activation.

git clone https://github.com/mangesh-ux/thinking-completion-probes.git
cd thinking-completion-probes
git checkout d71422e0f59fddcbff8960c876201301b8be4b4c
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install -e . -r requirements/cpu.txt huggingface_hub==1.19.0
hf download mangesh-ux/thinking-completion-probes cpu_reproduce.py \
  --repo-type dataset --revision v1.0.0 --local-dir hf-example
python hf-example/cpu_reproduce.py --output runs/hf-verify-v1

By default, cpu_reproduce.py anonymously downloads only the five numeric bundle files, restores saved models and verifies 39 historical statistics and paired predictions. It does not fit models, download upstream model weights or dataset text, run inference, collect test data or provision resources. Output is write-once; failures and partial downloads are retained. Choose a fresh output. Optional --refit runs the unchanged 640 grouped-CV plus five final TRAIN fits:

python hf-example/cpu_reproduce.py --output runs/hf-refit-v1 --refit

Privacy, license and citation

No questions, reference answers, reasoning text, raw prompt/generated token IDs, original problem IDs, source row indices, model/tokenizer files, private archives, cloud logs, secrets, machine paths or unused test problems are released. No identity lookup table is distributed. Anonymous group ordering preserves seeded bootstrap behavior. Activations/features may encode source information: anonymization is not formal privacy or a guarantee of non-invertibility.

Our code, documentation and generated numerical derivatives are MIT. This does not relicense third-party data or weights. DeepMath-103K's pinned dataset card declares MIT; consult its upstream-source acknowledgements and applicable terms. Qwen3-1.7B's pinned model card declares Apache-2.0. Those revision pins apply to dataset and model/tokenizer, respectively; no current main substitution is implied.

Cite CITATION.cff, DeepMath-103K (He et al., arXiv:2504.11456), and Qwen3 (Qwen team, arXiv:2505.09388):

Mangesh (mangesh-ux). Thinking Completion Probes: Exploratory Qwen3 Activation Forecasting, version 1.0.0 (2026). https://github.com/mangesh-ux/thinking-completion-probes