Datasets:

mvishiu11
/

nlp-shap-text-validation

text stringlengths 21 91	answer stringclasses 2 values	label_set listlengths 2 2
work as well as a remarkably faithful one	positive	[ "negative", "positive" ]
becomes a tv episode rather than a documentary	negative	[ "negative", "positive" ]
quite good at providing some good old fashioned spooks	positive	[ "negative", "positive" ]
is a beautiful aching sadness to it all	positive	[ "negative", "positive" ]
that recalls the classics of early italian neorealism	positive	[ "negative", "positive" ]
far the worst movie of the year	negative	[ "negative", "positive" ]
first-time director denzel washington and a top-notch cast	positive	[ "negative", "positive" ]
you thought would leave you cold	negative	[ "negative", "positive" ]
to pamela s emotional roller coaster life	positive	[ "negative", "positive" ]
sleepwalk through vulgarities in a sequel you can refuse	negative	[ "negative", "positive" ]
rises above the level of a telanovela	positive	[ "negative", "positive" ]
that kate is n't very bright but	negative	[ "negative", "positive" ]
the elements that will grab you	positive	[ "negative", "positive" ]
has not spawned a single good film	negative	[ "negative", "positive" ]
are capable of anteing up some movie star charisma	positive	[ "negative", "positive" ]
control-alt-delete simone as quickly as possible	negative	[ "negative", "positive" ]
a harrowing account of a psychological breakdown	positive	[ "negative", "positive" ]
laughs at how clever it s being	positive	[ "negative", "positive" ]
a funny and touching film that is gorgeously	positive	[ "negative", "positive" ]
that such a talented director as chen kaige	positive	[ "negative", "positive" ]
it s really unclear why this project was undertaken	negative	[ "negative", "positive" ]
that s surprisingly short of both adventure and song	negative	[ "negative", "positive" ]
this movie is careless and unfocused	negative	[ "negative", "positive" ]
their parents wise folks that they are read books	positive	[ "negative", "positive" ]
what s with all the shooting	negative	[ "negative", "positive" ]
the effect comes off as self-parody	negative	[ "negative", "positive" ]
i simply ca n't recommend it enough	positive	[ "negative", "positive" ]
got a house full of tots do n't worry	positive	[ "negative", "positive" ]
is one that should be thrown back in the river	negative	[ "negative", "positive" ]
this is a very funny heartwarming film	positive	[ "negative", "positive" ]
the human story is pushed to one side	negative	[ "negative", "positive" ]
skins is heartfelt and achingly real	positive	[ "negative", "positive" ]
as warm as it is wise	positive	[ "negative", "positive" ]
pretty and gifted it really won my heart	positive	[ "negative", "positive" ]
of an especially well-executed television movie	positive	[ "negative", "positive" ]
half the fun of college football games	positive	[ "negative", "positive" ]
grounded in an undeniable social realism	positive	[ "negative", "positive" ]
too campy to work as straight drama	negative	[ "negative", "positive" ]
rumor a muddled drama about coming to terms with death	negative	[ "negative", "positive" ]
is just an overexposed waste of film	negative	[ "negative", "positive" ]
a welcome if downbeat missive from a forgotten front	positive	[ "negative", "positive" ]
s the worst movie i ve seen this summer	negative	[ "negative", "positive" ]
an absolute delight for all audiences	positive	[ "negative", "positive" ]
set the women s liberation movement back 20 years	negative	[ "negative", "positive" ]
buried somewhere inside its fabric but never clearly seen	negative	[ "negative", "positive" ]
there is n't much there here	negative	[ "negative", "positive" ]
for its success on a patient viewer	positive	[ "negative", "positive" ]
have all the suspense of a 20-car pileup	negative	[ "negative", "positive" ]
directs an equally miserable film the following year	negative	[ "negative", "positive" ]
a movie where story is almost an afterthought	negative	[ "negative", "positive" ]
none of this is meaningful or memorable	negative	[ "negative", "positive" ]
of jim brown a celebrated wonder in the spotlight	positive	[ "negative", "positive" ]
s neither too erotic nor very thrilling	negative	[ "negative", "positive" ]
were hailed as the works of an artist	positive	[ "negative", "positive" ]
loud and offensive but more often	negative	[ "negative", "positive" ]
to the genre and another first-rate performance	positive	[ "negative", "positive" ]
frequently veers into corny sentimentality probably	negative	[ "negative", "positive" ]
plodding soap opera disguised as a feature film	negative	[ "negative", "positive" ]
with his fake backdrops and stately pacing	negative	[ "negative", "positive" ]
sleepwalk through vulgarities in a sequel you can refuse	negative	[ "negative", "positive" ]
the movie winds up feeling like a great missed opportunity	negative	[ "negative", "positive" ]
you ll still have a good time	positive	[ "negative", "positive" ]
the title many can aspire but none can equal	negative	[ "negative", "positive" ]
on and on to the point of nausea	negative	[ "negative", "positive" ]
whine the bellyaching of a paranoid and unlikable man	negative	[ "negative", "positive" ]
one of the best movies of the year	positive	[ "negative", "positive" ]
almost do n't notice the 129-minute running time	positive	[ "negative", "positive" ]
the comedy equivalent of saddam hussein	negative	[ "negative", "positive" ]
the one not-so-small problem with expecting	negative	[ "negative", "positive" ]
is throwing up his hands in surrender	negative	[ "negative", "positive" ]
every scrap is of the darkest variety	positive	[ "negative", "positive" ]
sucks but has a funny moment or two	negative	[ "negative", "positive" ]
its loving yet unforgivingly inconsistent depiction	positive	[ "negative", "positive" ]
a great premise but only a great premise	negative	[ "negative", "positive" ]
that has no point and goes nowhere	negative	[ "negative", "positive" ]
meticulously uncovers a trail of outrageous force and craven concealment	positive	[ "negative", "positive" ]
seems embarrassed by his own invention	negative	[ "negative", "positive" ]
a beer with but they re simply not funny performers	negative	[ "negative", "positive" ]
a nice little story in the process	positive	[ "negative", "positive" ]
opera on film is never satisfactory	negative	[ "negative", "positive" ]
just a string of stale gags	negative	[ "negative", "positive" ]
does n't deserve a passing grade even on a curve	negative	[ "negative", "positive" ]
never rises to its clever what-if concept	negative	[ "negative", "positive" ]
director lee has a true cinematic knack	positive	[ "negative", "positive" ]
is one of the year s best films	positive	[ "negative", "positive" ]
squandering his opportunity to make absurdist observations	negative	[ "negative", "positive" ]
one of the year s best films	positive	[ "negative", "positive" ]
can see the forest for the trees	positive	[ "negative", "positive" ]
is plenty of fun for all	positive	[ "negative", "positive" ]
just might turn on many people to opera in general	positive	[ "negative", "positive" ]
has all the trappings of an energetic extreme-sports adventure but	positive	[ "negative", "positive" ]
the last kiss is really all about performances	positive	[ "negative", "positive" ]
rude scarily funny sorrowfully sympathetic to the damage it surveys	positive	[ "negative", "positive" ]
in which this fascinating and timely content comes wrapped	positive	[ "negative", "positive" ]
a formula comedy redeemed by its stars	positive	[ "negative", "positive" ]
predisposed to the movie s rude and crude humor	negative	[ "negative", "positive" ]
be pleasant in spite of its predictability	positive	[ "negative", "positive" ]
that as a director washington demands and receives excellent performances	positive	[ "negative", "positive" ]
lifts the film high above run-of-the-filth gangster flicks	positive	[ "negative", "positive" ]
marry science fiction with film noir and action flicks	positive	[ "negative", "positive" ]

End of preview. Expand in Data Studio

NLP Shapley — Text Validation & Faithfulness Data

Consolidated experimental data for the study "Is Shapley attribution on LLMs faithful, and if so, in what specific way?" This repo backs the analysis and the paper write-up; it is a living dataset — some experiments are still running on the cluster and will be added here as they land (see Status below).

The research arc (what this data answers)

Machinery is correct. Exact vs sampled Shapley converge; efficiency holds (campaign/).
The logprob game is NOT faithful on open-ended prompts. Removing the top-k Shapley tokens is statistically indistinguishable from removing random tokens. Reason: the game's target moves with the coalition (v(S) = Σ logprob of the text the model regenerates under S) — it rewards fluency, not decision content. Evidence: faithfulness_logprob/, games/.
Fix the target. target_prob (logP(fixed answer O | S)) removes the moving target; games/ compares it against logprob.
The useful game: contrastive decision game. On closed-answer tasks, v(S) = logP(correct O | S) − logP(foil F | S) cancels the fluency baseline. Usefulness claim (demonstrated with a held-out discrete flip check): the tokens Shapley ranks highest are the minimal set whose removal flips the model's decision (and whose presence restores it), beating LOO / LIME / integrated gradients / attention / random. Evidence: decision_game/, ground_truth/.

Layout

Path	What it is
`campaign/`	Consolidated core-campaign analysis: estimator convergence, cross-model, value-fn sensitivity (CSV), `report.md`, `EXECUTIVE_SUMMARY.md`, plots.
`attributions.parquet`	Consolidated per-token Shapley attributions from the core campaign (see `ANALYSIS.md`).
`faithfulness_logprob/`	`baseline__<model>.parquet` + `.summary.json`: SV vs LOO/LIME/IG/attention on a held-out teacher-forced logprob of the original answer; absence-policy sweep; Wilcoxon + BH-FDR.
`games/`	`game__<model>.parquet` + `.summary.json`: exact SV on logprob vs target_prob from one cached 2ⁿ sweep; P(O)-drop comprehensiveness + behavioural divergence vs random; Spearman(φ_logprob, φ_target).
`ground_truth/<jobid>/`	Planted-trigger fine-tune + exact-SV top-1 recovery of the trigger, base vs fine-tuned. `sv_recovery_{base,ft}.parquet`, `summary.json`.
`decision_game/<task>/<jobid>/`	Contrastive decision game usefulness: `decision__<task>__<model>.parquet` + `.summary.json`. Metrics: `min_flip_set`, `flip@k`, `restore@k` per method; synthetic trigger recovery. Tasks: `synthetic` (planted trigger), `sentiment` (SST-2).
`data/`	Prepared inputs: `sent_{train,eval}.jsonl` (short, balanced SST-2, punctuation-stripped so words == players).
`ANALYSIS.md`	Narrative analysis of the core campaign attributions.

Key schemas

decision_game parquet rows: idx, n, task, method, correct_full, min_flip_set, flippable, k, flip, restore (+ trigger_top1, trigger_mass, flipset_is_trigger for the synthetic task). method ∈ {shapley, loo, lime, integrated_grad, attention, random}. min_flip_set = smallest #top-φ tokens whose removal flips the discrete predicted label (smaller = better).
decision_game summary.json: per_method (mean/median min-flip-set, flip@k, restore@k, trigger recovery), shapley_vs_baselines (paired Wilcoxon, p_sv_smaller, win-rate), fdr_bh_qvalues.
games summary.json: per game (logprob, target_prob) → comp_target and beh_div with mean_top, mean_random, gap_top_minus_random, winrate_top_gt_random, p_top_gt_random; plus spearman_phi_lp_vs_tp_mean.
ground_truth summary.json: base/ft → task_accuracy, and per game (logprob_game, target_prob_game) → sv_top1_recovery, sv_mass_trigger_mean.

Models

Attribution/faithfulness: Qwen/Qwen2.5-3B-Instruct, google/gemma-2-2b-it, meta-llama/Llama-3.2-3B-Instruct. Controlled fine-tune tasks: Qwen/Qwen2.5-0.5B-Instruct.

Status (2026-09-02)

Present now: core campaign, one faithfulness baseline (qwen25_3b), a historical ground-truth run, the sentiment inputs, and the decision-game smoke. Still running / to be added: remaining baseline + game models, dual-game ground-truth, and the full decision-game runs (synthetic, sentiment). This card will be refreshed as they complete.

Reproduction

Harness, scripts, and a full runbook live in the nlp-shap-research repository (scripts/{game_faithfulness,ground_truth,decision_game,baseline_faithfulness}.py, docs/plans/infrastructure/Experiment Harness & Helios Runbook.md).

Downloads last month: 26