model-organisms-for-real/ao-training-task-test-splits

Activation Oracle training-task test splits

Held-out classification test splits for the Activation Oracle (AO) training tasks, generated for ten base models that Adam Karvonen released oracles for but never published training data for (hf_push_to_hub=False in his run configs).

Each .pt file contains the test split for one (dataset, variant, base model) triple, with precomputed activation vectors — the AO pipeline forces save_acts=True for test splits, so these are ready to evaluate without re-running the base model.

Contents

200 files: 10 base models × 10 classification datasets × 2 variants.

Datasets: geometry_of_truth, relations, sst2, md_gender, snli, ag_news, ner, tense, language_identification, singular_plural.

Two variants per dataset:

• —single — max_window_size=1, num_qa_per_sample=2: exactly one activation vector is injected and the oracle answers yes/no from it.
• —multi — max_window_size=50, num_qa_per_sample=1: a contiguous span of k ~ U[1, 50] vectors ending a few tokens from the end of the sequence.

Both variants are included because the upstream build_datasets keys eval data by dataset_config.dataset_name, which is identical for the two loaders, so the multi loader overwrites the single one and in-training eval only ever scored multi-token items.

Generation config

All splits use layer_percents = [25, 50, 75] and hook_onto_layer = 1, matching every oracle in Karvonen's "Activation Oracles" collection (read from each repo's ao_config.json). The classification loader params match his exactly: min_end_offset=-1, max_end_offset=-5, num_test=250, seed=42.

The test items are the same underlying sentences and questions for every base model: get_classification_datapoints carves the test set out as all_examples[-250:] after a seed-42 shuffle, and that carve-out does not depend on num_train. Only the activations, tokenization and sampled windows differ between models. This makes cross-model comparison clean.

Usage

Drop the files into sft_training_data/ in a checkout of activation_oracles; the config-hashed filenames are what the loaders look for, so they will be picked up as cache hits rather than regenerated:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="model-organisms-for-real/ao-training-task-test-splits",
    repo_type="dataset",
    local_dir="sft_training_data",
    allow_patterns=["*Qwen3-8B*"],   # or omit to fetch all ten models
)

Loading one file directly:

import torch
from nl_probes.utils.dataset_utils import TrainingDataPoint

obj = torch.load("classification_sst2_model_Qwen3-8B_n_250_save_acts_False_test_<hash>.pt")
data = [TrainingDataPoint(**d) for d in obj["data"]]
print(data[0].layer, data[0].steering_vectors.shape, data[0].target_output)

Note the save_acts_False in the filename reflects the config field; test splits always store activations regardless, so steering_vectors is populated.

Provenance and caveats

• —Generated with a fork of the upstream repo: nikxtaco/activation_oracles, branch raffaello/ao-training-task-evals, via experiments/eval_training_tasks.py.
• —Activations are collected from the clean base model, no adapter applied.
• —These files contain activation vectors derived from license-gated base models (Gemma, Llama). The corresponding model licenses govern derivative artifacts; check them before redistributing.
• —The equivalent splits for allenai/OLMo-2-0425-1B-SFT and google/gemma-3-1b-it are not here — they are already published in model-organisms-for-real/olmo2_1b_sft_checkpoint_oracle_v1-training-data and model-organisms-for-real/gemma3_1b_it_oracle_v1-training-data, and use four layer percents ([25, 50, 75, 88] and [25, 50, 75, 96]) rather than three.
• —meta-llama/Llama-3.3-70B-Instruct is absent: 282 GB of weights, and Karvonen's config quantized it to 8-bit, which would not be comparable to these bf16 runs.

Results

Eval results using these splits, across 13 oracles, are in the fork under experiments/AO_TRAINING_TASK_EVALS.md.