moviesentiment-multitask-onnx-int8

Model Card — MovieSentiment DistilBERT INT8

Following the Mitchell et al. 2019 Model Card template.

Model details

• Model name: moviesentiment-classifier
• Architecture: DistilBERT-base-uncased (66M parameters), HuggingFace distilbert-base-uncased
• Head: 2-class linear classifier on [CLS] pooled output
• Quantization: ONNX dynamic INT8 (8-bit weights, FP32 activations) via optimum.onnxruntime.ORTQuantizer with AutoQuantizationConfig.avx512_vnni(is_static=False)
• Size: 64 MB (down from 256 MB FP32; 4× compression)
• Framework versions: transformers>=4.46, optimum[onnxruntime]>=1.17, onnxruntime>=1.17
• Owner: Soumya Sarkar — github.com/Cryptic2-0
• License: MIT
• Date: 2026-05-26 (last retrained from main)
• Git SHA: returned at GET /version

Intended use

Primary use: classify English IMDb-style movie reviews into positive/negative sentiment.

Primary users: this project is a portfolio MLOps demonstration. The API is publicly reachable but is not intended for downstream products.

Out-of-scope:

• Non-English text — see Quantitative analyses below for the cliff-edge accuracy drop.
• Product, food, news, or any non-movie domain — not evaluated; results unreliable.
• Long-form text > 5000 characters — truncated at 512 tokens.
• Multi-class sentiment (neutral, mixed, sarcastic) — binary output only.
• Any safety-, health-, or finance-sensitive decision.

Factors

Metrics

Test set: held-out 5,000 reviews from IMDb 50K.

CI for F1 (bootstrap, 1000 resamples): [0.934, 0.944].

Quantitative analyses

Performance by review length

Model is weakest on very short reviews (terse one-liners with weak signal). This matches intuition — fewer tokens, fewer keywords.

Performance by confidence bucket

The model is well-calibrated above 0.85 — high-confidence predictions are reliable. The thin (~2%) low-confidence band is essentially chance-level and should be treated as "model declined to answer."

Sarcasm slice

Reviews flagged by the heuristic sarcasm seed list (~340 examples): F1 0.823, vs 0.939 overall — a 12-point absolute drop. This is the most prominent failure mode and is documented on the live demo card.

Non-English

Spanish translations of the test set (machine-translated via NLLB): F1 0.51 (chance is 0.50). Treat any non-English input as untrusted. The frontend does not gate on language; this is intentional (interview material), but a production deployment should.

Per-genre slice (live IMDb scrape, 2026-05-29)

Evaluated on a fresh live IMDb GraphQL scrape of the 10 movie IDs in params.yaml. 2,525 reviews retrieved, sorted by HELPFULNESS_SCORE (IMDb's default — biases toward positive reviews). Each movie's primary genre per IMDb's listing. Source data at data/processed/live_with_genre.parquet; raw numbers at metrics/per_genre_f1.json.

Overall: n=2,525, accuracy 0.942, macro F1 0.825.

Key caveat — class skew: the live test set is ~93% positive because IMDb's helpfulness-sort top-loads positive reviews. On the original balanced HF test set (5,000 examples, 50/50) macro F1 is 0.939; the live slice's lower macro F1 (0.825) reflects the imbalance, not a model regression. Binary F1 on the majority (positive) class stays in the 0.964–0.973 band across all genres, consistent with the original eval.

What the slice does and doesn't tell us: Action has the weakest macro F1 (0.779), which means the model is most likely to miss negative reviews of Action films. Adventure and Crime hold the strongest macro F1, but those buckets also have less class skew. A genre-balanced re-scrape — sort by date instead of helpfulness — would be the next step to disentangle "genre effect" from "class-balance effect". Documented as a follow-up in docs/future_improvements.md.

Ethical considerations

• Training data is movie-only. Generalization to other "subjective text" domains is not implied.
• No demographic factors are inferred or stored. The pipeline does not capture reviewer identity.
• Reservoir-sampled production inputs are stored under data/production/recent.parquet for drift detection. The current sampler is in-memory and flushed every 100 inserts; users who post personally identifying text are subject to this storage for the lifetime of the running container.
• No human-in-the-loop. Predictions are returned directly.

Caveats and recommendations

1. Use confidence to gate downstream actions — anything below 0.70 should be surfaced as "uncertain", not classified.
2. The model was trained on IMDb 50K which is balanced (50/50 positive/negative). Real-world review distributions skew positive (~70/30). Expect more false negatives at deployment time.
3. If you need higher recall on negative reviews (e.g., for moderation), retrain with a re-weighted loss or threshold-tune via the bundled eval/metrics.py PR-curve helpers.
4. Retraining loop closes when Evidently drift > 0.30 (monitor/drift.py) — fires a GH Actions workflow to scrape, retrain (SageMaker spot + LoRA), re-export ONNX, redeploy.

Environmental impact

Trained once on ml.g4dn.xlarge (T4 GPU) for ~25 minutes. Estimated energy: ~0.05 kWh. With CodeCarbon's default Australia grid factor (~0.66 kgCO2/kWh) that's ~33g CO2 per training run. The retraining loop runs at most weekly.

Inference: ARM Graviton Fargate, 0.25 vCPU, 1 GB RAM, single task. Steady-state ~3 W. Per-prediction CO2 cost is negligible at portfolio traffic.

Where to find more

• Source: github.com/Cryptic2-0/Project
• Live API contract: OpenAPI /docs on the running Fargate task
• Drift reports: docs/drift_reports/ (HTML, generated by python -m moviesentiment.cli drift)
• Latency benchmarks: docs/benchmarks.md
• Load test report: docs/load_test_report.html
• Datasheet for the training data: docs/datasheet.md