icl-heads/language-generalisation

Language Generalisation

The 8-shot evidence-tracking set machine-translated into Chinese, Hindi, Spanish, French and Arabic (NLLB-200-3.3B), plus the English source and a Chinese variant that keeps the English Yes/No labels.

Part of the icl-heads collection: the datasets behind a study of which Llama-3.1-8B-Instruct attention heads mediate in-context evidence accumulation and in-context label mapping, discovered with Differentiable Circuit Masking (a learned per-head mask over K/V activations patched at the demonstration positions).

Layout

One subdirectory per dataset variant; split filenames are preserved as they are used by the training and evaluation code.

narcissism_8_now_answer_this_ar/
  contrastive_source.csv
  contrastive_test.csv
  contrastive_train.csv
narcissism_8_now_answer_this_en/
  contrastive_source.csv
  contrastive_test.csv
  contrastive_train.csv
narcissism_8_now_answer_this_es/
  contrastive_source.csv
  contrastive_test.csv
  contrastive_train.csv
narcissism_8_now_answer_this_fr/
  contrastive_source.csv
  contrastive_test.csv
  contrastive_train.csv
narcissism_8_now_answer_this_hi/
  contrastive_source.csv
  contrastive_test.csv
  contrastive_train.csv
narcissism_8_now_answer_this_zh/
  contrastive_source.csv
  contrastive_test.csv
  contrastive_train.csv
narcissism_8_now_answer_this_zh_yesno/
  contrastive_source.csv
  contrastive_test.csv
  contrastive_train.csv

Row format

Each row is a contrastive pair. prompt_context and counter_context hold the demonstrations for the two forward passes; unlabeled_input is the shared final question; prompt_gold and counter_gold are the answers the unpatched model gives on each. Where the answer vocabulary is replaced, yes_replacement and no_replacement name the two label words for that row, and the training loss is a Jensen-Shannon divergence restricted to those tokens and renormalized.

Provenance

Personas derive from Anthropic's model-written evaluations. Generated and used by the belief_dynamics research repository; see its reports/ directory for the experiments these datasets support.