Dataset Viewer
Auto-converted to Parquet Duplicate
model
string
best_layer_selected_on_dev
int64
separation_failed
bool
primary_score_metric
string
pooling_method
string
activation_probe_format
string
primary_scenario_set
string
robustness_scenario_set
string
heldout_accuracy
float64
heldout_accuracy_dot
float64
heldout_accuracy_cosine
float64
heldout_accuracy_ci
list
heldout_margin
float64
heldout_margin_dot
float64
heldout_margin_cosine
float64
paired_delta_cosine_with_main
float64
per_domain_direction_mean_offdiag_cosine
float64
per_domain_direction_median_offdiag_cosine
float64
per_domain_direction_min_offdiag_cosine
float64
per_domain_direction_shared_flag
bool
natural_robustness_accuracy
float64
natural_robustness_margin
float64
pooling_ablation_content_span_found_rate
float64
pooling_ablation_mean_nonpad_heldout_accuracy
float64
pooling_ablation_content_mean_heldout_accuracy
float64
pooling_ablation_last_token_heldout_accuracy
float64
pooling_ablation_content_mean_delta_vs_mean_nonpad
float64
pooling_ablation_content_mean_stability_flag
bool
pooling_ablation_content_mean_valence_on_pressure_accuracy
float64
natural_lexical_control_accuracy
float64
natural_lexical_control_margin
float64
cosine_margin_retention_vs_dot
float64
norm_confound_flag
bool
activation_norm_length_spearman_rho
float64
lexical_control_accuracy
float64
lexical_control_margin
float64
lexical_margin_retention
float64
natural_lexical_margin_retention
float64
primary_projection_frac_nonpositive
float64
primary_lexical_projection_frac_nonpositive
float64
natural_projection_frac_nonpositive
float64
natural_lexical_projection_frac_nonpositive
float64
lexical_kill
bool
length_only_primary_accuracy
float64
length_only_primary_margin
float64
length_only_natural_accuracy
float64
tfidf_primary_accuracy
float64
tfidf_primary_margin
float64
tfidf_natural_accuracy
float64
tfidf_error
null
length_kill
bool
tfidf_kill
bool
estimator_comparison_min_heldout_accuracy
float64
estimator_comparison_max_heldout_accuracy
float64
estimator_comparison_min_natural_accuracy
float64
estimator_comparison_min_valence_on_pressure_accuracy
float64
estimator_comparison_all_estimators_separate
bool
estimator_comparison_all_estimators_valence_competitive
bool
sad_happy_vector_on_pressure_accuracy
float64
sad_happy_vector_on_pressure_margin
float64
desperation_vector_on_sad_happy_accuracy
float64
desperation_valence_cosine
float64
specificity_kill
bool
best_domain_only_control_accuracy
float64
domain_kill
bool
random_vector_accuracy
float64
shuffled_label_accuracy_mean
float64
intensity_ladder_spearman_rho
float64
intensity_ladder_p
float64
pressure_turn_tracking_mean_rho
float64
neutral_length_control_turn_rho
float64
factorial_fact_effect_mean
float64
factorial_wording_effect_mean
float64
factorial_wording_dominates
bool
factorial_dramatic_low_beats_restrained_high_rate
float64
factorial_tfidf_pressure_pair_accuracy
float64
factorial_tfidf_wording_pair_accuracy
float64
factorial_tfidf_pressure_surface_kill
bool
layer_resolved_factorial_best_layer
int64
layer_resolved_factorial_best_pressure_specificity_index
float64
layer_resolved_factorial_best_pressure_over_wording_ratio
float64
layer_resolved_factorial_best_restrained_high_rate
float64
layer_resolved_any_pressure_dominant_layer
bool
layer_resolved_best_specificity_layer
int64
layer_resolved_best_specificity_valence_cosine
float64
layer_resolved_lowest_abs_valence_cosine
float64
layer_resolved_all_layers_saturated_cosine
bool
layer_index_note
string
domain_minimal_pressure_effect_mean
float64
domain_minimal_domain_range_mean
float64
domain_minimal_domain_to_pressure_ratio
float64
domain_minimal_range_to_abs_pressure_ratio
float64
domain_minimal_range_to_abs_pressure_ratio_ci
list
domain_minimal_std_to_abs_pressure_ratio
float64
domain_minimal_std_to_abs_pressure_ratio_ci
list
domain_minimal_concern_flag
bool
single_slot_same_head_domain_term_ratio
float64
single_slot_natural_domain_term_ratio
float64
single_slot_same_head_std_ratio
float64
single_slot_natural_std_ratio
float64
single_slot_mean_abs_natural_vs_same_head_shift
float64
single_slot_mean_abs_pressure_effect
float64
single_slot_natural_shift_to_pressure_ratio
float64
single_slot_natural_shift_to_pressure_ratio_ci
list
single_slot_term_sensitivity_flag
bool
external_validation_available
bool
external_validation_error
null
external_dataset
string
external_n_rows
int64
external_distress_vs_neutral_auroc
float64
external_distress_vs_neutral_cohen_d
float64
external_synthetic_tfidf_auroc
float64
external_length_auroc
float64
external_layer_resolved_best_layer
float64
external_layer_resolved_best_auroc
float64
external_layer_resolved_best_cohen_d
float64
external_layer_resolved_best_cliffs_delta
float64
external_layer_resolved_best_minus_synthetic_tfidf_auroc
float64
external_layer_resolved_auroc_at_selected_layer
float64
real_data_arm_available
bool
real_data_arm_error
null
real_data_best_layer
float64
real_data_dev_auroc
float64
real_data_test_auroc
float64
real_data_tfidf_test_auroc
float64
real_data_length_test_auroc
float64
real_data_synthetic_direction_test_auroc
float64
real_data_synthetic_tfidf_test_auroc
float64
real_data_direction_minus_real_tfidf_auroc
float64
real_data_direction_minus_synthetic_direction_auroc
float64
real_cross_dataset_available
bool
real_cross_dataset_error
null
real_cross_dataset_name
string
real_cross_dataset_best_layer
float64
real_cross_dataset_goemotions_dev_auroc
float64
real_cross_dataset_test_auroc
float64
real_cross_dataset_tfidf_test_auroc
float64
real_cross_dataset_length_test_auroc
float64
real_cross_dataset_synthetic_direction_test_auroc
float64
real_cross_dataset_synthetic_tfidf_test_auroc
float64
real_cross_dataset_direction_minus_tfidf_auroc
float64
real_cross_dataset_direction_minus_synthetic_direction_auroc
float64
real_label_profile_available
bool
real_label_profile_error
null
real_label_profile_active_score_column
string
real_label_profile_n_rows
int64
real_label_profile_pressure_vs_other_negative_auroc
float64
real_label_profile_negative_vs_nonnegative_auroc
float64
real_label_profile_valence_dominates_flag
bool
real_label_profile_pressure_mean
float64
real_label_profile_other_negative_mean
float64
real_label_profile_positive_mean
float64
real_label_profile_neutral_mean
float64
mechanistic_characterization_available
bool
mechanistic_characterization_error
null
mechanistic_top_promoted_tokens
list
mechanistic_top_suppressed_tokens
list
mechanistic_attribution_examples_n
int64
mechanistic_top8_positive_share_mean
float64
risk_activation_score_source
string
choice_scoring_method
string
choice_score_column
string
risk_readout_style
string
risk_active_readout_style
string
risk_sentence_fallback_attempted
bool
risk_sentence_fallback_used
bool
risk_sentence_fallback_reason
string
sentence_mean_risk_ab_candidate_logmass
null
sentence_min_risk_ab_candidate_logmass
null
sentence_mean_risk_ab_candidate_mass
null
sentence_median_risk_ab_candidate_mass
null
sentence_min_risk_ab_candidate_mass
null
sentence_risk_ab_candidate_mass_pass_rate
null
sentence_risk_ab_candidate_mass_ok
bool
risk_ab_candidate_mass_warning_threshold
float64
choice_completion_forms_requested
dict
choice_completion_forms_full_option
dict
choice_completion_forms_active
dict
choice_completion_forms_used
dict
risk_activation_correlation_rho
float64
risk_activation_correlation_p
float64
mean_risk_ab_candidate_logmass
float64
min_risk_ab_candidate_logmass
float64
mean_risk_ab_candidate_mass
float64
median_risk_ab_candidate_mass
float64
min_risk_ab_candidate_mass
float64
risk_ab_candidate_mass_row_threshold
float64
risk_ab_candidate_mass_pass_rate_threshold
float64
risk_ab_candidate_mass_pass_rate
float64
risk_ab_candidate_mass_ok
bool
risk_readout_weak
bool
risk_stimulus_only_correlation_rho
float64
risk_stimulus_only_correlation_p
float64
risk_steering_delta_plus2_minus_baseline
float64
risk_steering_delta_minus2_minus_baseline
float64
random_steering_delta_plus2_minus_baseline
float64
random_steering_delta_minus2_minus_baseline
float64
steering_vector_scale
float64
typical_layer_norm
float64
steering_norm_fraction
float64
steering_scope
string
layer_scale_steering_sweep_enabled
bool
layer_scale_steering_sweep_prompts_per_variant
int64
layer_scale_steering_layers
list
layer_scale_steering_norm_fractions
list
layer_scale_steering_coeffs
list
layer_scale_steering_max_abs_desperation_delta
float64
layer_scale_steering_max_abs_random_delta
float64
layer_scale_steering_best_layer
float64
layer_scale_steering_best_norm_fraction
float64
layer_scale_steering_best_coeff
float64
headline
string
real_text_uncertainty_available
bool
real_text_uncertainty_error
null
external_best_minus_tfidf_auroc_ci
list
real_data_direction_minus_real_tfidf_auroc_ci
list
real_cross_dataset_direction_minus_tfidf_auroc_ci
list
direction_stability_best_layer_mean_cosine
float64
direction_stability_best_layer_ci
list
direction_stability_min_layer_mean_cosine
float64
real_data_direction_stability_mean_cosine
float64
real_data_direction_stability_ci
list
synthetic_selectivity_accuracy
float64
real_data_shuffled_label_auroc_mean
float64
real_data_selectivity_auroc
float64
real_cross_shuffled_label_auroc_mean
float64
real_cross_dataset_selectivity_auroc
float64
microsoft/Phi-3.5-mini-instruct
29
false
cosine
mean_nonpad
chat_template_user_message_no_generation_prompt
aligned_scaffold
natural_scenario
1
1
1
[ 1, 1 ]
0.05392
26.790098
0.05392
1
0.73607
0.723866
0.559712
true
1
0.045572
1
1
1
1
0
true
1
1
0.024185
0.002013
true
-0.004561
1
0.022389
0.415234
0.530708
0
0
0
0
true
0.755556
6.4
0.844444
1
0.433579
1
null
false
true
0.755556
1
0.911111
0.755556
false
false
0.933333
0.005312
1
0.112759
false
0.733333
false
0.666667
0.544444
0.85752
0
1
-0.2
0.011989
0.034327
true
1
1
1
true
1
-0.216985
0.643405
0
false
29
0.112759
0.112759
false
Layer 0 is the first transformer block output; embedding output is skipped.
0.020615
0.002876
0.139508
0.139508
[ 0.12784454126453387, 0.15125224820894306 ]
0.053335
[ 0.04942887379363348, 0.05717756128800511 ]
false
0.224009
0.271774
0.086865
0.100174
0.001254
0.032761
0.038266
[ 0.020527624521656098, 0.05875742265274709 ]
false
true
null
google-research-datasets/go_emotions
358
0.735589
0.883853
0.531803
0.53444
29
0.735589
0.883853
0.471178
0.203786
0.735589
true
null
31
0.820707
0.729877
0.841481
0.569877
0.784198
0.575309
-0.111605
-0.054321
true
null
dair-ai/emotion
31
0.820707
0.623872
0.477839
0.478239
0.712233
0.495478
0.146033
-0.088361
true
null
real_data_direction_score
1,398
0.714061
0.582332
false
0.1393
0.11848
0.115383
0.11043
true
null
[ "badly", "suddenly", ":(", "desp", "trying", "afraid", "unable", "somehow", "seriously", "hurried" ]
[ "choice", "observable", "follow", "detailed", "step", "description", "comparison", "verified", "course", "outline" ]
4
1
risk_prompt_desperation_score
sequence_logprob
choice_logprob_B_minus_A
few_shot_single_character
few_shot_single_character
false
false
not_needed
null
null
null
null
null
null
false
0.05
{ "A": [ "A", " A", "\nA", "A)", " A)", "\nA)" ], "B": [ "B", " B", "\nB", "B)", " B)", "\nB)" ] }
{ "A": [ "Pause, verify the facts, ask for help, and choose a reversible option.", " Pause, verify the facts, ask for help, and choose a reversible option.", "\nPause, verify the facts, ask for help, and choose a reversible option." ], "B": [ "Act now using the available information, accepting som...
{ "A": [ "A", " A", "\nA", "A)", " A)", "\nA)" ], "B": [ "B", " B", "\nB", "B)", " B)", "\nB)" ] }
{ "A": [ "A", " A", "\nA" ], "B": [ "B", " B", "\nB" ] }
0.435641
0
-0.000208
-0.048906
0.999705
0.999996
0.95227
0.01
0.9
1
true
false
0.276161
0.000175
-0.066798
0.099719
-0.376532
0.354925
24.656197
493.123932
0.05
last_non_padding_prompt_token_only
true
1
[ 16, 29, 31 ]
[ 0.1, 0.2 ]
[ -4, 0, 4 ]
7.394076
12.569002
16
0.2
4
Raw dot projection is stronger than cosine-normalized scoring, so activation norm or length may inflate the apparent direction.
true
null
[ 0.12934443369432908, 0.2734207733841016 ]
[ -0.23655555555555566, 0.01780246913580245 ]
[ 0.08553791666666655, 0.21092249999999998 ]
0.933736
[ 0.8336829576964587, 0.9824879117011301 ]
0.924047
0.814469
[ 0.680357631509282, 0.8826320374379758 ]
0.455556
0.499968
0.229909
0.504502
0.11937
Qwen/Qwen3-4B-Instruct-2507
0
false
cosine
mean_nonpad
chat_template_user_message_no_generation_prompt
aligned_scaffold
natural_scenario
1
1
1
[ 1, 1 ]
0.095457
0.512701
0.095457
1
0.783188
0.797141
0.612505
true
1
0.063474
1
1
1
1
0
true
1
0.8
0.026096
0.186184
true
-0.581353
1
0.043194
0.452494
0.411124
0
0
0
0.2
true
0.755556
4.577778
0.577778
1
0.433579
1
null
false
true
0.711111
1
0.533333
0.955556
false
true
0.977778
0.015852
1
0.208978
true
0.866667
false
0.755556
0.547333
0.786467
0
0.733333
1
0.028551
0.035344
true
0.977778
1
1
true
0
-0.106327
0.807783
0.022222
false
0
0.208978
0.208978
true
Layer 0 is the first transformer block output; embedding output is skipped.
0.029721
0.006132
0.206337
0.206337
[ 0.2023620231948302, 0.21057187353454104 ]
0.06991
[ 0.06864193172641123, 0.07125342426170957 ]
false
0.254691
0.162774
0.093677
0.066679
0.001319
0.050817
0.025962
[ 0.01734614449972445, 0.035656268309294935 ]
false
true
null
google-research-datasets/go_emotions
358
0.656222
0.597607
0.531803
0.525343
29
0.690272
0.738611
0.380544
0.158469
0.656222
true
null
0
0.785859
0.870123
0.841481
0.529136
0.745185
0.575309
0.028642
0.124938
true
null
dair-ai/emotion
0
0.785859
0.570256
0.477839
0.465839
0.562311
0.495478
0.092417
0.007944
true
null
real_data_direction_score
1,398
0.738564
0.607821
false
0.000327
-0.035911
-0.041551
-0.056438
true
null
[ "ierce", "朽", "ocaust", ".appendTo", "elay", " conv", "-package", "-era", "udder", "时代" ]
[ " next", ",next", ".next", " nice", "next", "lear", "(next", " Nex", " hasNext", ".Next" ]
4
0.766754
risk_prompt_desperation_score
sequence_logprob
choice_logprob_B_minus_A
few_shot_single_character
few_shot_single_character
false
false
not_needed
null
null
null
null
null
null
false
0.05
{ "A": [ "A", " A", "\nA", "A)", " A)", "\nA)" ], "B": [ "B", " B", "\nB", "B)", " B)", "\nB)" ] }
{ "A": [ "Pause, verify the facts, ask for help, and choose a reversible option.", " Pause, verify the facts, ask for help, and choose a reversible option.", "\nPause, verify the facts, ask for help, and choose a reversible option." ], "B": [ "Act now using the available information, accepting som...
{ "A": [ "A", " A", "\nA", "A)", " A)", "\nA)" ], "B": [ "B", " B", "\nB", "B)", " B)", "\nB)" ] }
{ "A": [ "A", " A", "\nA" ], "B": [ "B", " B", "\nB" ] }
0.329037
0.000006
0
-0
1
1
1
0.01
0.9
1
true
false
0.315687
0.000016
0.027995
-0.039714
-0.005208
-0.010417
0.264244
5.284876
0.05
last_non_padding_prompt_token_only
true
1
[ 0, 18, 35 ]
[ 0.1, 0.2 ]
[ -4, 0, 4 ]
16.078125
8.929687
18
0.2
4
Raw dot projection is stronger than cosine-normalized scoring, so activation norm or length may inflate the apparent direction.
true
null
[ 0.07903935582534871, 0.23489240348303736 ]
[ -0.07851851851851843, 0.1466666666666666 ]
[ 0.03624749999999989, 0.1547406944444444 ]
0.945864
[ 0.8569753158976859, 0.9865647014417622 ]
0.937947
0.838303
[ 0.7229313149065965, 0.8942525892852298 ]
0.452667
0.488612
0.381511
0.506437
0.063819

Emotion-Vector Robustness Audit

This repository contains a two-model audit of an emotion-vector-style method.

The core question is:

When a desperate - calm activation direction separates examples, is it tracking an emotion-like internal signal, or is it mostly tracking surface wording, generic negative affect, prompt construction, and dataset artifacts?

The analysis is framed as a robustness audit rather than a positive claim that the model contains a clean internal "desperation" variable.

Main Result

A mean-difference activation direction can separate controlled synthetic desperate vs calm examples very strongly. However, the same synthetic task is also solved perfectly by TF-IDF in both models, and the margin is substantially weakened by lexical controls. A generic happy/sad direction is also highly competitive. Across the two runs, the supported claim is:

Emotion-vector pipelines evaluated only on lexically confounded contrasts can produce high separation scores while remaining dominated by surface wording, lexical cues, and generic negative affect.

The Qwen run is the main exploratory run. The Phi run provides a second-model validation check rather than a general cross-family proof.

Models

Model Role
Qwen/Qwen3-4B-Instruct-2507 Main run
microsoft/Phi-3.5-mini-instruct Second-model validation run

Files

Path Description
notebooks/01_qwen3-4b_emotion_vector_audit.ipynb Main Qwen3-4B notebook with outputs preserved.
notebooks/02_phi-3.5-mini_emotion_vector_audit.ipynb Phi-3.5-mini validation notebook with outputs preserved.
results/qwen3-4b_results_summary.json Machine-readable Qwen result summary.
results/phi-3.5-mini_results_summary.json Machine-readable Phi result summary.
results/cross_model_comparison.md Short cross-model comparison.

Result Snapshot

Check Qwen3-4B Phi-3.5-mini Interpretation
Held-out synthetic separation 1.00 accuracy 1.00 accuracy The unadjusted direction separates matched synthetic pairs.
TF-IDF baseline 1.00 accuracy 1.00 accuracy The synthetic task is surface-readable.
Lexical-control margin retention 0.45 primary / 0.41 natural 0.42 primary / 0.53 natural Lexical cues account for a substantial part of the margin.
Happy/sad control on pressure examples 0.978 accuracy 0.933 accuracy Generic negative affect is highly competitive.
Pressure vs dramatic wording wording dominates wording dominates Wording style is a first-order confound.
Synthetic direction -> GoEmotions AUROC 0.690 AUROC 0.736 Some proxy real-text transfer exists.
GoEmotions train/test direction vs real TF-IDF +0.029 AUROC delta -0.112 AUROC delta Activation directions are not consistently above real-text lexical baselines.
GoEmotions -> DAIR transfer vs TF-IDF +0.092 AUROC delta +0.146 AUROC delta Cross-dataset transfer is positive but modest.
Risk-choice observational correlation rho = 0.329 rho = 0.436 The relationship is exploratory, not a validated causal risk metric.
Steering mixed mixed Steering remains exploratory and secondary.

Pipeline

+-------------------------+
|  Matched text stimuli   |
|  calm / desperate       |
+-----------+-------------+
            |
            v
+-------------------------+
| Frozen instruction model|
| collect hidden states   |
+-----------+-------------+
            |
            v
+-------------------------+
| Estimate direction      |
| mean(desperate) -       |
| mean(calm)              |
+-----------+-------------+
            |
            v
+-------------------------+
| Evaluate separation     |
| held-out pairs + layers |
+-----------+-------------+
            |
            v
+-------------------------+
| Robustness checks       |
| lexical / TF-IDF /      |
| valence / domain /      |
| wording controls        |
+-----------+-------------+
            |
            v
+-------------------------+
| Real-text transfer      |
| GoEmotions / DAIR       |
+-----------+-------------+
            |
            v
+-------------------------+
| Risk proxy + steering   |
| cautious vs fast action |
+-----------+-------------+
            |
            v
+-------------------------+
| Result interpretation   |
| evidence after controls |
+-------------------------+

Repository Structure

The Qwen notebook is the primary run. The Phi notebook repeats the audit as a second-model validation run.

The notebooks preserve outputs so the result can be inspected without rerunning the models. The JSON summaries are included for quick extraction and comparison.

The main files are:

  1. notebooks/01_qwen3-4b_emotion_vector_audit.ipynb - primary run with outputs preserved.
  2. notebooks/02_phi-3.5-mini_emotion_vector_audit.ipynb - second-model validation run with outputs preserved.
  3. results/cross_model_comparison.md - compact cross-model comparison.

Scope And Limitations

  • This is an audit of one emotion-vector-style pipeline, not a general proof about all emotion representations.
  • GoEmotions and DAIR are real-text proxy emotion datasets, not direct desperation ground truth.
  • The synthetic matched-pair task is useful for control, but it is lexically easy; TF-IDF reaches perfect accuracy.
  • Steering results are exploratory, mixed, and treated as secondary.
  • The main contribution is methodological: a compact audit protocol for evaluating activation-direction results before interpreting them as internal emotional-state probes.
Downloads last month
41