Taste Lab
Music · Film · PoliticsQwen 2.5 7B
Vector trained on
1. Music preference
200 works · 100 pairsMean choice probabilityCritical acclaim .Commercial success .
Not runCritical acclaim .Commercial success .
Does this music preference carry over to film?
The vector was trained on music only. No film or political examples were used. The same music-trained vector is applied below.
2. Film transfer
200 works · 100 pairsMean choice probabilityCritical acclaim .Commercial success .
Not runCritical acclaim .Commercial success .
Does this music preference carry over to LeftValues questions?
The vector was trained on music only. No film or political examples were used. The same music-trained vector is applied below.
3. Political preference
LeftValues · 72 questionsLeftValues axesNot run
Revolution .Reform .
Scientific .Utopian .
Centralization .Decentralization .
Internationalism .Nationalism .
Party .Union .
Industry .Ecology .
Conservative .Progressive .
AgreeDisagree
Method & data sources
- Models
- Qwen 2.5 7B Instruct and Meta Llama 3.1 8B Instruct. Each model and training domain uses its own vector. Vector values are never shared across models or training domains.
- Vector training
- 32 pairs from 2023 Pitchfork and IFPI lists train broad A/B music preferences. Music-only validation selects the vector. Base model weights are frozen.
- Music grid
- 100 critical-list songs from Pitchfork 2021 and 2020, paired with 100 Billboard Year-End Hot 100 songs from 2021. Each song appears once; all are disjoint from training.
- Film grid
- 100 critical-list films and 100 commercial-list films, paired once per work. Each group includes 80 films from Jee & Shaw and 20 additional films from TSPDT or Box Office Mojo.
- Politics grid
- 72 original LeftValues English statements with the original five answer options. Scores are probability-weighted model expectations over the five options, not human discrete answers.
- Politics score
- Each axis score is computed from the original LeftValues question weights. The model sees each statement once with original and reversed option order, without axis labels.
- Answer format
- Art comparisons use two-choice constrained JSON. Politics constrains answer JSON and derives five-option probabilities from the permitted option logits.
- Steering fit
- The music direction is trained on broad A/B music preference choices; the Qwen music direction keeps a 0.25 multiplier selected on music validation.
- Calibrated range
- -1 to +1, evaluated on the selected training domain only. Stronger settings may become non-monotonic.
- Animation
- Sequential reveal of completed model measurements
- Interpretation
- Lists overlap in cultural meaning. Artists, genres, geography, eras and model familiarity are not fully matched. This is not a causal test of human habitus. LeftValues measures divisions within left political traditions. This use is an exploratory model measurement, not a validated survey result.