Oxford-HIPlab/iclr2026-lm-logprobs

LM Log-Probabilities for Value Bias Analysis

Next-token log-probability distributions from 12 language models across 54 prompts, used in the paper:

Reward Models Inherit Value Biases from Pretraining

Brian Christian, Jessica A.F. Thompson, Elle, Vincent Adam, Hannah Rose Kirk, Christopher Summerfield, Tsvetomira Dumbalska (ICLR 2026)

Part of the Oxford-HIPlab collection for this paper.

Dataset description

Each CSV contains the full next-token log-probability distribution (log-softmax of last-token logits) for one language model evaluated on 54 prompts. These prompts follow a factorial design: 6 adjectives (best, greatest, good, worst, terrible, bad) x 3 superlatives (ever, in the world, of all time) x 3 concision styles (in one word, in a single word, please answer in one word only).

Columns

Column	Description
token_id	Vocabulary index
token_name	Raw token string from the tokenizer
token_decoded	Decoded token (human-readable)
best_ever_one, best_ever_single, ...	Log-probability (54 prompt columns)

Each row is one token from the model's vocabulary. Gemma models have ~256K tokens; Llama models have ~128K tokens.

Files

File	Model	Type	Family	Size
google--gemma-2-2b.csv	google/gemma-2-2b	Pretrained	Gemma	277 MB
google--gemma-2-2b-it.csv	google/gemma-2-2b-it	Instruction-tuned	Gemma	274 MB
google--gemma-2-9b-it.csv	google/gemma-2-9b-it	Instruction-tuned	Gemma	273 MB
google--gemma-2-27b-it.csv	google/gemma-2-27b-it	Instruction-tuned	Gemma	275 MB
meta-llama--Llama-3.2-3B.csv	meta-llama/Llama-3.2-3B	Pretrained	Llama	138 MB
meta-llama--Llama-3.2-3B-Instruct.csv	meta-llama/Llama-3.2-3B-Instruct	Instruction-tuned	Llama	139 MB
meta-llama--Llama-3.2-1B-Instruct.csv	meta-llama/Llama-3.2-1B-Instruct	Instruction-tuned	Llama	139 MB
meta-llama--Meta-Llama-3-8B-Instruct.csv	meta-llama/Meta-Llama-3-8B-Instruct	Instruction-tuned	Llama	140 MB
meta-llama--Llama-3.1-8B-Instruct.csv	meta-llama/Llama-3.1-8B-Instruct	Instruction-tuned	Llama	139 MB
meta-llama--Meta-Llama-3-70B-Instruct.csv	meta-llama/Meta-Llama-3-70B-Instruct	Instruction-tuned	Llama	139 MB
meta-llama--Llama-3.1-70B-Instruct.csv	meta-llama/Llama-3.1-70B-Instruct	Instruction-tuned	Llama	139 MB
meta-llama--Llama-3.3-70B-Instruct.csv	meta-llama/Llama-3.3-70B-Instruct	Instruction-tuned	Llama	138 MB

Total size: ~2.2 GB

Usage

Quick download (Python)

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Oxford-HIPlab/iclr2026-lm-logprobs",
    filename="google--gemma-2-2b.csv",
    repo_type="dataset",
)

With the paper's code

Clone the code repository and run:

pip install -r requirements.txt
python scripts/download_data.py          # downloads all 12 CSVs into data/logprobs/
python figures/generate_figure_2.py      # reproduce Figure 2

Load a single file with pandas

import pandas as pd

df = pd.read_csv("google--gemma-2-2b.csv")
logprobs = df["greatest_ever_one"]            # log-probs for one prompt
probs = logprobs.apply(lambda x: 2.718**x)    # convert to probabilities

Generation details

Log-probabilities were generated using scripts/generate_logprobs.py from the code repository. For each model and prompt:

1. The prompt is tokenized (using apply_chat_template for instruction-tuned models, plain tokenization for pretrained models).
2. A single forward pass produces logits at the final token position.
3. log_softmax is applied to obtain log-probabilities over the full vocabulary.

All computations use bfloat16 precision with deterministic settings (torch.use_deterministic_algorithms(True), seed 42).

Citation

@inproceedings{christian2026reward,
  title={Reward Models Inherit Value Biases from Pretraining},
  author={Christian, Brian and Thompson, Jessica A F and Elle and Adam, Vincent and Kirk, Hannah Rose and Summerfield, Christopher and Dumbalska, Tsvetomira},
  booktitle={International Conference on Learning Representations},
  year={2026}
}