THE ONE SUCCESSFUL RUN -- output, pasted exactly as Kaggle printed it ==================================================================== Date: 16 October 2026 (late). Platform: Kaggle, free GPU, NVIDIA Tesla T4. Notebook: a fresh notebook, created to get away from a stuck session. Script that produced this: kaggle_one_cell_fixed.py (in this folder) Transformers library: 5.16.1 Graphics chip available: True ====================================================================== STEP 1 of 6 -- reading the data ====================================================================== FOUND learn.jsonl FOUND check.jsonl FOUND test_inputs.jsonl FOUND test_answers.tsv Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads. reviews to learn from : 22500 reviews to check with : 2500 exam reviews : 24877 exam answers : 24877 checks passed ====================================================================== STEP 2 of 6 -- loading the starting model (downloads about 270 MB) ====================================================================== model loaded onto: cuda [transformers] DistilBertForSequenceClassification LOAD REPORT from: distilbert-base-uncased - UNEXPECTED: vocab_layer_norm.*, vocab_transform.*, vocab_projector.bias (not needed, ignored) - MISSING: classifier.weight, classifier.bias, pre_classifier.* (new, random -- this is why the model starts at ~50%) ====================================================================== STEP 3 of 6 -- measuring the model BEFORE training (takes a few minutes) ====================================================================== untrained model, exam accuracy : 50.5% always answer 'neg', exam accuracy : 49.9% About half means no skill yet, which is the correct starting point. ====================================================================== STEP 4 of 6 -- TRAINING. This is the long part. ====================================================================== reading the 22,500 reviews 3 times, 16 at a time a check score will be printed after each pass preparing the reviews (a couple of minutes)... ready [2112/2112 16:22, Epoch 3/3] Epoch Training Loss Validation Loss Accuracy 1 0.500789 0.444729 0.913600 2 0.330639 0.458983 0.916400 3 0.236240 0.518115 0.920800 ====================================================================== STEP 5 of 6 -- scoring the trained model on the exam ====================================================================== BEFORE training : 50.5% AFTER training : 91.3% good reviews : 11,520 of 12,472 called correctly bad reviews : 11,185 of 12,405 called correctly Published results for models this size on this dataset sit in the low 90s. ====================================================================== FIVE IT GOT WRONG ====================================================================== TRUE NEG MODEL SAID POS "Two wildlife photographers are traveling through the Amazon River basin on their latest assignment. While trying to capture the wildlife of the area on film, our photographers cross paths with a game hunter, who is stalking the animals for another reason. Looking to eliminate th... TRUE POS MODEL SAID NEG Dan Burgess is a nice guy. He happens to be a Christian. Dan can't get a date with a girl and thinks that all of his friends are having all of the fun. He is constantly being bothered by non-believers and being made fun of.

Dan prays one night, and wishes he was never ... TRUE NEG MODEL SAID POS What Game Over: Kasparov and the Machine does best is to delve into Garry Kasparov's psyche during the 1997 competition against IBM's Deep Blue. You see him becoming more and more paranoid, and increasingly unravelled, all because in the second game, Deep Blue made a move that se... TRUE POS MODEL SAID NEG Barry Kane (Robert Cummings) is wrongfully accused on sabotaging a hanger making aircrafts for the war. He goes on the run, meets Patricia Martin (Priscilla Lane) along the way, and she joins him to find and bring the real criminals to justice.

There are a lot of thing... TRUE POS MODEL SAID NEG Otto Preminger was one of the great maverick film directors.Like John Huston he was a character from one of his own movies,like Mr Huston he was a wonderful ham who slipped in and out of performance at random but the final product of his labours was unmistakably his own. With a f... ====================================================================== FIVE IT GOT RIGHT ====================================================================== TRUE POS MODEL SAID POS TV churns out dozens of true-crime movies every year. You can see 3 or 4 every Saturday on Lifetime, and Court TV can be relied on for a few every weekend. So I started watching The Morrison Murders thinking I knew very well what to expect: a more or less competent retelling of a... TRUE NEG MODEL SAID NEG 20 years ago Dr Chopper(yes, he is driving a chopper as well as chopping people up) disappeared when the police found dead bodies at his clinic. Since then people have been disappearing around Lake Tanoka(or something) and dead bodies/parts of bodies has been found all over the p... TRUE POS MODEL SAID POS I can't imagine why it hasn't been theatrically released yet. It's got a great ensemble cast, with Sutherland, Lane, and especially Chris Evans doing spectacular work. Wake up, studio execs!

The story is based upon the experiences of the author/screenwriter, growing up... TRUE POS MODEL SAID POS A unique blend of musical, film-noir and comedy - with a few sex scenes thrown in for good measure. The only other film I can think of with a fairly similarly wild and madcap mixture of themes and clichés is the French movie Billy Ze Kick - but that has a more surreal and quirky ... TRUE POS MODEL SAID POS Gilles Mamouni is playing with the audience with the story of Max (Vincent Cassel) in search of his biggest lost love (Monica Bellucci) just before to get engaged to another woman. Mamouni uses many flashbacks sequences without warning so the best way to know where we are in the ... ====================================================================== STEP 6 of 6 -- saving the model ====================================================================== saved to: /kaggle/working/movie-review-sorter 0.0 MB config.json 267.8 MB model.safetensors 0.7 MB tokenizer.json 0.0 MB tokenizer_config.json NOTES ON THIS OUTPUT - The run took about 30 minutes in total; 16 minutes 22 seconds of it was the training itself (5.5 minutes per pass). - The exam was 24,877 reviews that the model never saw during training. - 11,520 + 11,185 = 22,705 of 24,877 = 91.27%. - Both answers were learned well: 92.4% of good reviews, 90.2% of bad reviews. A model that had just learned "always say positive" would show 50% on one row. - Two harmless warnings appeared during the run: * "unauthenticated requests to the HF Hub" -- a download rate-limit notice. * "Was asked to gather along dimension 0..." -- an internal note from PyTorch. - The paste above stopped at tokenizer_config.json; the final line "DONE." printed just below it.