sonar-sae / night5_json /w10_decoder_reading.json
nickypro's picture
Upload folder using huggingface_hub
7c2241d verified
Raw
History Blame Contribute Delete
20.6 kB
{
"meta": {
"experiment": "W10_decoder_reading",
"n_sentences": 500,
"decoder_card": "text_sonar_basic_decoder",
"n_decoder_layers": 24,
"n_cross_attn_heads": 16,
"z_is_single_position": true,
"skeptic_note": "z is ONE vector => decoder cross-attends to a single position (softmax over 1 key ~ degenerate). Localist test reframed as z-DIRECTION ablation -> output-token-change vs matched-norm random control."
},
"A_layerwise_emergence": {
"n_sentences_probed": 300,
"n_content_tokens": 4032,
"n_surface_tokens": 4413,
"median_emerge_frac_content": 0.6086956521739131,
"median_emerge_frac_surface": 1.0,
"mean_emerge_frac_content": 0.6372390441683919,
"mean_emerge_frac_surface": 0.8954768027271205,
"note": "emerge_frac = layer index (0..1) at which lens-top1 first locks to the final token and stays. lower=earlier.",
"logitlens_agree_content_by_layer": [
0.0,
0.0,
0.000496031746031746,
0.000496031746031746,
0.001240079365079365,
0.004712301587301587,
0.012152777777777778,
0.023809523809523808,
0.04290674603174603,
0.0689484126984127,
0.13392857142857142,
0.2708333333333333,
0.39211309523809523,
0.5354662698412699,
0.6495535714285714,
0.7090773809523809,
0.7584325396825397,
0.8038194444444444,
0.8353174603174603,
0.8648313492063492,
0.8953373015873016,
0.890625,
0.8871527777777778,
0.9995039682539683
],
"logitlens_agree_surface_by_layer": [
0.0,
0.0,
0.0,
0.0,
0.005211874008610922,
0.012236573759347382,
0.015862225243598458,
0.019034670292318152,
0.02605937004305461,
0.04713346929526399,
0.05755721731248584,
0.10786313165646952,
0.12735100838431906,
0.22818944029005211,
0.28030818037616134,
0.4049399501472921,
0.37049626104690686,
0.41536369816451396,
0.4421028778608656,
0.4579651031044641,
0.4400634489009744,
0.27781554498073874,
0.35871289372309084,
0.9435757987763427
]
},
"B_crossattn_read_by_layer": {
"note": "mean ||cross_attn_output|| / ||residual_into_crossattn|| per layer; how strongly z is read at each depth.",
"xattn_read_frac_by_layer": [
11.981749674479167,
5.466305143038432,
7.206465433438619,
9.383789989153543,
10.489324464797974,
9.529227778116862,
10.383447993596395,
11.11037944952647,
7.5069301795959475,
10.289530377388001,
10.700671984354654,
10.48009269396464,
11.221453116734823,
12.288406060536703,
8.880138290723165,
8.605154465039572,
9.007333512306213,
8.356222885449727,
8.236391728719076,
5.922921441396078,
5.301507139205933,
3.4041185712814332,
1.5054391316572826,
1.156544069846471
],
"peak_read_layer": 13,
"early_third_mean": 9.443836240768434,
"mid_third_mean": 9.996547146042188,
"late_third_mean": 5.361309809982776
},
"A_lexical_bag_first": {
"note": "frac of final content-token IDs already present ANYWHERE in lens-top1 (bag) by layer.",
"content_bag_recall_by_layer": [
0.994312718671078,
0.9931090149673741,
0.9920685508019101,
0.9905864079447673,
0.9803983530229664,
0.950794700182858,
0.9084043334403238,
0.8453719021489914,
0.7591730988023548,
0.6281410301085218,
0.5171946775274948,
0.49037106638345035,
0.5285413721125798,
0.6213316098675999,
0.7044672625419526,
0.7454325638814795,
0.780488143005944,
0.8166848795482686,
0.8439021312221768,
0.8739997629506138,
0.9017810559753284,
0.8959827667172865,
0.8926178860126225,
0.9992361111111111
],
"bag_recall_at_half_depth": 0.5285413721125798
},
"C_zdirection_influence": {
"note": "Perturb z along a direction, re-greedy-decode, measure #words changed (symmetric diff vs baseline word set) + Jaccard. Localist => content-word-ablation dirs change FEW, targeted words (the ablated word vanishes) more than a matched-norm random perturbation does.",
"perturbation_magnitude": "content_ablate=full word removal from pool; pca/rand=0.5*||z|| along dir",
"content_ablate_dir": {
"n": 359,
"mean_words_changed": 1.1114206128133706,
"std": 1.6467602832006853,
"mean_jaccard_to_base": 0.9415911581469185,
"frac_targeted_word_vanished": 0.2785515320334262
},
"random_dir_control": {
"n": 240,
"mean_words_changed": 2.4625,
"std": 3.0397796658091303,
"mean_jaccard_to_base": 0.8827266463037415
},
"pca_dirs": {
"pca0": {
"mean_words_changed": 11.9,
"std": 6.859786196862601,
"mean_jaccard_to_base": 0.5001729309860259
},
"pca1": {
"mean_words_changed": 6.666666666666667,
"std": 5.377938473264845,
"mean_jaccard_to_base": 0.7162827654127722
},
"pca2": {
"mean_words_changed": 9.45,
"std": 6.103072996450231,
"mean_jaccard_to_base": 0.5888467391794264
},
"pca3": {
"mean_words_changed": 9.716666666666667,
"std": 6.258571899580784,
"mean_jaccard_to_base": 0.6079678794291621
},
"pca4": {
"mean_words_changed": 10.283333333333333,
"std": 6.537307464766277,
"mean_jaccard_to_base": 0.5825735608825627
},
"pca5": {
"mean_words_changed": 8.716666666666667,
"std": 5.413229678810567,
"mean_jaccard_to_base": 0.6322222442879579
}
},
"localist_signal": "content-ablation changes fewer/targeted words than random",
"examples": [
{
"sentence": "New Bedford, MA - A local man was arrested and charged with heroin distribution after a multi-agency investigation, according to authorities.",
"baseline_decode": "New Bedford, MA - A local man was arrested and charged with distributing heroin after a multi-agency investigation, according to authorities.",
"examples": [
{
"dir": "New",
"kind": "content_ablate",
"decode": "New Bedford, MA - A local man was arrested and charged with distributing heroin after a multi-agency investigation, according to authorities."
},
{
"dir": "Bedford",
"kind": "content_ablate",
"decode": "New Bedford, MA - A local man was arrested and charged with distributing heroin after a multi-agency investigation, according to authorities."
},
{
"dir": "local",
"kind": "content_ablate",
"decode": "New Bedford, MA - A man was arrested and charged with distributing heroin after a multi-agency investigation, according to authorities."
},
{
"dir": "man",
"kind": "content_ablate",
"decode": "New Bedford, MA - A local man was arrested and charged with distributing heroin after a multi-agency investigation, according to authorities."
},
{
"dir": "arrested",
"kind": "content_ablate",
"decode": "New Bedford, MA - A local man was arrested and charged with distribution of heroin after a multi-agency investigation, according to authorities."
},
{
"dir": "charged",
"kind": "content_ablate",
"decode": "New Bedford, MA - A local man was arrested and charged with distribution of heroin after a multi-agency investigation, according to authorities."
},
{
"dir": "pca0",
"kind": "pca",
"decode": "New Bedford, MA - And our own people are in a position to be confronted by the police and the police."
},
{
"dir": "pca1",
"kind": "pca",
"decode": "New Bedford, MAYBROW - A local man was arrested and charged with a heroine after a multi-agent investigation, according to the authorities."
}
]
},
{
"sentence": "The school's annual awards ceremony saw students praised for their academic, artistic, and sporting achievements.",
"baseline_decode": "The school's annual awards ceremony saw students being praised for their academic, artistic, and sports achievements.",
"examples": [
{
"dir": "school",
"kind": "content_ablate",
"decode": "The school's annual awards ceremony saw students being praised for their academic, artistic, and sports achievements."
},
{
"dir": "annual",
"kind": "content_ablate",
"decode": "The school's awards ceremony saw students being praised for their academic, artistic, and sports achievements."
},
{
"dir": "awards",
"kind": "content_ablate",
"decode": "The school's annual award ceremony saw students being praised for their academic, artistic, and sports achievements."
},
{
"dir": "ceremony",
"kind": "content_ablate",
"decode": "The school's annual awards ceremony saw students being praised for their academic, artistic, and sports achievements."
},
{
"dir": "saw",
"kind": "content_ablate",
"decode": "The school's annual awards ceremony saw students being praised for their academic, artistic, and sports achievements."
},
{
"dir": "students",
"kind": "content_ablate",
"decode": "The school's annual awards ceremony saw students being praised for their academic, artistic, and sports achievements."
},
{
"dir": "pca0",
"kind": "pca",
"decode": "The school's annual awards ceremony sees students praised for their achievements and achievements."
},
{
"dir": "pca1",
"kind": "pca",
"decode": "The school's annual awards feast at school, students were praised for their academic, artistic, and sports achievements."
}
]
},
{
"sentence": "As I ascended to the rooftop Gravity Bar, I was treated to stunning views of the city and a unique perspective on the brewing process.",
"baseline_decode": "As I climbed up to the rooftop Gravity Bar, I was given a view of the city's amazing and a unique perspective on the brewing process.",
"examples": [
{
"dir": "ascended",
"kind": "content_ablate",
"decode": "As I climbed on the rooftop Gravity Bar, I was given a breathtaking view of the city and a unique perspective on the brewing process."
},
{
"dir": "rooftop",
"kind": "content_ablate",
"decode": "As I climbed to the Gravity Bar, I was given a view of the city's amazing and a unique perspective on the brewing process."
},
{
"dir": "Gravity",
"kind": "content_ablate",
"decode": "As I climbed up to the rooftop Bar of Grace, I was given a view of the city's amazing and a unique perspective on the brewing process."
},
{
"dir": "Bar",
"kind": "content_ablate",
"decode": "As I climbed to the rooftop Gravity Bar, I was given a breathtaking view of the city and a unique perspective on the brewing process."
},
{
"dir": "treated",
"kind": "content_ablate",
"decode": "As I climbed up on the rooftop Gravity Bar, I was given a breathtaking view of the city and a unique perspective on the brewing process."
},
{
"dir": "stunning",
"kind": "content_ablate",
"decode": "As I climbed up to the rooftop Gravity Bar, I was given a view of the city and a unique perspective on the brewing process."
},
{
"dir": "pca0",
"kind": "pca",
"decode": "As I get on to the surface on gravity bar, I get to see the beautiful view of the world and an unusual understanding of the process of brewing."
},
{
"dir": "pca1",
"kind": "pca",
"decode": "As I climbed up on the rooftop Gravity Bar, I was graced with the breathtaking view of the city and a unique look at the groove of the brew."
}
]
},
{
"sentence": "I'm happy to report that some of the best cookies I've ever had were from beloved institutions like Levain Bakery and Jacques Torres Chocolate.",
"baseline_decode": "I am happy to report that some of the best cookies I have ever had were from beloved institutions like Levain Bakery and Jacques Torres Chocolate.",
"examples": [
{
"dir": "happy",
"kind": "content_ablate",
"decode": "I am happy to report that some of the best cookies I have ever had were from beloved institutions like Levain Bakery and Jacques Torres Chocolate."
},
{
"dir": "report",
"kind": "content_ablate",
"decode": "I'm happy to report that some of the best cookies I've ever had were from beloved institutions like Levain Bakery and Jacques Torres Chocolate."
},
{
"dir": "some",
"kind": "content_ablate",
"decode": "I am happy to report that some of the best cookies I have ever had were from beloved institutions like Levain Bakery and Jacques Torres Chocolate."
},
{
"dir": "best",
"kind": "content_ablate",
"decode": "I am happy to report that some of the best cookies I have ever had were from beloved institutions like Levain Bakery and Jacques Torres Chocolate."
},
{
"dir": "cookies",
"kind": "content_ablate",
"decode": "I am happy to report that some of the best cakes I have ever had were from beloved institutions like Levain Bakery and Jacques Torres Chocolate."
},
{
"dir": "ever",
"kind": "content_ablate",
"decode": "I am happy to report that some of the best cookies I've had were from beloved institutions like Levain Bakery and Jacques Torres Chocolate."
},
{
"dir": "pca0",
"kind": "pca",
"decode": "I'm happy to tell you that some of the best memories I've ever had are from my parents and my friends."
},
{
"dir": "pca1",
"kind": "pca",
"decode": "I'm happy to tell you some of the finest chocolates I've ever had were from the lovely Levain Bakery and Jacques Torres Chocolate."
}
]
},
{
"sentence": "Under the deal, Saudi Arabia and Iran will increase their combined oil production to 12 million barrels per day, from 9 million currently.",
"baseline_decode": "Under the deal, Saudi Arabia and Iran will increase their combined oil production to 12 million barrels per day, from 9 million currently.",
"examples": [
{
"dir": "Under",
"kind": "content_ablate",
"decode": "Under the deal, Saudi Arabia and Iran will increase their combined oil production to 12 million barrels per day, from 9 million currently."
},
{
"dir": "deal",
"kind": "content_ablate",
"decode": "Under the deal, Saudi Arabia and Iran will increase their combined oil production to 12 million barrels per day, from 9 million currently."
},
{
"dir": "Saudi",
"kind": "content_ablate",
"decode": "Under the deal, Saudi Arabia and Iran will increase their combined oil production to 12 million barrels per day, from 9 million currently."
},
{
"dir": "Arabia",
"kind": "content_ablate",
"decode": "Under the deal, Saudi Arabia and Iran will increase their combined oil production to 12 million barrels per day, from 9 million currently."
},
{
"dir": "Iran",
"kind": "content_ablate",
"decode": "Under the deal, Saudi Arabia and Iran will increase their combined oil production to 12 million barrels per day, from 9 million currently."
},
{
"dir": "increase",
"kind": "content_ablate",
"decode": "Under the deal, Saudi Arabia and Iran will increase their combined oil production to 12 million barrels per day, from 9 million currently."
},
{
"dir": "pca0",
"kind": "pca",
"decode": "Under the deal, we will increase our oil production and oil production to 9 billion barrels a day."
},
{
"dir": "pca1",
"kind": "pca",
"decode": "Under the deal, Saudia Arabia and Iran and her combined oil production will rise to 12 million barrels a day, from 9 million barrels today."
}
]
},
{
"sentence": "Perched atop the rolling hills of the Scottish Highlands, Edinburgh is a city that seamlessly blends its rich history with modern architectural splendor.",
"baseline_decode": "Edinburgh is situated on the hillside of the Scottish Highlands, an city that seamlessly blends its rich history with modern architectural splendor.",
"examples": [
{
"dir": "Perched",
"kind": "content_ablate",
"decode": "Edinburgh is built on the rolling hills of the Scottish Highlands, an city that seamlessly mixes its rich history with modern architectural splendor."
},
{
"dir": "atop",
"kind": "content_ablate",
"decode": "Edinburgh is situated on the rolling hills of the Scottish Highlands, an city that seamlessly blends its rich history with modern architectural splendor."
},
{
"dir": "rolling",
"kind": "content_ablate",
"decode": "Edinburgh is situated on the hillsides of the Scottish Highlands, an city that seamlessly blends its rich history with modern architectural splendor."
},
{
"dir": "hills",
"kind": "content_ablate",
"decode": "Edinburgh is situated on the rocky slopes of the Scottish Highlands, an city that seamlessly blends its rich history with modern architectural splendor."
},
{
"dir": "Scottish",
"kind": "content_ablate",
"decode": "Situated on the crest of the rolling hills of the Edinburgh Highlands, Edinburgh is a city that seamlessly blends its rich history with modern architectural splendor."
},
{
"dir": "Highlands",
"kind": "content_ablate",
"decode": "Edinburgh is situated on the crest of the rolling hills of Scottish Highlands, an city that seamlessly mixes its rich history with modern architectural splendor."
},
{
"dir": "pca0",
"kind": "pca",
"decode": "Is it not true that the present-day world is a mess of uncertainties and uncertainties?"
},
{
"dir": "pca1",
"kind": "pca",
"decode": "Edinburgh is nestled on the hillside of the rocky slopes of the Scottish Highlands, Edinburgh is a town that unmistakably mixes its dazzling history with its splendor of modern architecture."
}
]
}
]
},
"SCORING": {
"prediction": "decoder reads z DIFFUSELY (single position, no per-output-token z-direction map); reconstructs content/topic EARLY, refines surface form LATE; reading is DISTRIBUTED (no clean token->z-dir map).",
"falsifier": "clean per-output-token z-direction decode map (localist handle) OR z merely sets an initial state (trivial reading).",
"single_position_confirmed": true,
"content_emerges_before_surface": true,
"z_read_across_depth_not_just_init": true,
"clean_localist_token_to_zdir_map": false,
"verdict": "CONFIRMED: diffuse/distributed reading; content early, surface late"
}
}