Detecting Hallucinations in Large Language Models Using Semantic Entropy

This repository contains the code necessary to reproduce the short-phrase and sentence-length experiments of the Nature submission "Detecting Hallucinations in Large Language Models Using Semantic Entropy."

About This Fork

This is a fork of the original repository with modifications to enable execution on consumer-grade GPUs and additional features. Extensive documentation of the changes and adaptations is planned.

Added Arguments

--second_gpu

• Enables sharding across secondary low-power GPUs asymmetrically.
• Note: VRAM limits are currently hardcoded (0, 6GB, 6GB). A new argument to configure these limits is planned.

--metric=custom_llm

• Allows the use of a custom LLM for computing accuracy (e.g., Llama 3.2).

--custom_metric_model_name

• Complements the --metric=custom_llm argument. Specify the model name here.

--alt_entail_prompt

• Uses an alternative entailment prompt for use in context-specific entailment testing (customer service here). Works only with custom entailment LLMs.

Running Independent Computations

To compute uncertainty measures independently, use the following command:

python compute_uncertainty_measures.py \
    --metric=custom_llm \
    --custom_metric_model_name=Llama-3.2-3B-Instruct \
    --entailment_model=Llama-3.2-3B-Instruct \
    --no-compute_accuracy_at_all_temps \
    --eval_wandb_runid=sddy15no \
    --entity=theflyingrahul-indian-institute-of-management-bangalore \
    --restore_entity_eval=theflyingrahul-indian-institute-of-management-bangalore

Fixes and Additions from the Original Repository

1. Whitespace Handling:

   • Resolved an issue where additional whitespaces in LLM responses caused a ValueError. This was hotpatched by approximating the output window shift.

2. Model and Dataset Support:

   • Added support for additional models and datasets. (TODO: Document this in detail.)

3. Demonstrations:

   • Included demonstrations for:
     • Dissecting the generated .pkl files.
     • Analyzing stochasticity in outputs.

Install and configure accelerate/DeepSpeed for training

Accelerate should be installed in conda? Check if it works

Use the latest version of hf-transformers

Gemma 3 support is flaky. The visual stack in Gemma 3 (which is multi-modal) is unnecessary for our work. Text-only support is hotpatched in the latest version of hf-transformers.

• Install as: pip install --upgrade git+https://github.com/huggingface/transformers.git
• See: huggingface/transformers#36683
• See: huggingface/transformers#36741

Hotfixes

• Upgraded packages for Gemma 3 support
  • Running torch v2.6.x on CUDA/NVCC 12.8 (instead of 11.8 from the Nature article).
  • Running hf-transformers v4.52.x.

Other notes

• Build flash_attn:
  • MAKEFLAGS="-j$(nproc)" pip install flash_attn --no-build-isolation --use-pep517 --no-cache-dir
  • Flash Attention is a little stupid. Sometimes it works, sometimes it doesn't! Especially on Lyra. Need to debug this further.
• Some blocks to accommodate CECI Lyra HPCC (like $GLOBALSCRATCH) will be present: need to refactor this and add it as a command-line argument.
• Second GPU implementation is incomplete/not-verified for Gemma models since moved to Lyra.

TODOs

• Document the added model and dataset support.
• Add an argument to configure VRAM limits for --second_gpu.
• Implement DeepSpeed support.
• Implement Flash Attention 2 support.
• Configure hf-accelerate.