generate-responses

Generate model responses for benchmark tasks, optionally with steering applied. This command is useful for creating response datasets that can be evaluated later.

Basic Usage
python -m wisent generate-responses MODEL --task TASK --output FILE [OPTIONS]

Examples

Basic Response Generation
python -m wisent generate-responses meta-llama/Llama-3.1-8B-Instruct \
  --task truthfulqa_mc1 \
  --num-questions 50 \
  --output ./responses/truthfulqa_baseline.json
With Steering
python -m wisent generate-responses meta-llama/Llama-3.1-8B-Instruct \
  --task arc_easy \
  --num-questions 100 \
  --use-steering \
  --steering-object ./vectors/accuracy.pt \
  --steering-strength 1.5 \
  --output ./responses/arc_steered.json
Custom Generation Parameters
python -m wisent generate-responses meta-llama/Llama-3.1-8B-Instruct \
  --task gsm8k \
  --num-questions 20 \
  --max-new-tokens 256 \
  --temperature 0.3 \
  --top-p 0.9 \
  --verbose \
  --output ./responses/gsm8k.json

Arguments

Required

ArgumentDescription
modelModel name or path
--taskTask name (e.g., arc_easy, truthfulqa_mc1)
--outputOutput file path for results

Generation Options

ArgumentDefaultDescription
--num-questions10Number of questions to generate responses for
--max-new-tokens128Maximum tokens to generate
--temperature0.7Temperature for generation
--top-p0.95Top-p for nucleus sampling
--deviceautoDevice to use (cpu, cuda, mps)

Steering Options

ArgumentDefaultDescription
--use-steeringfalseUse steering during generation
--steering-object-Path to steering object file (.pt)
--steering-strength1.0Steering strength multiplier
--disable-thinkingfalseDisable thinking mode (Qwen models)

Related Commands

Stay in the loop. Never miss out.

Subscribe to our newsletter and unlock Wisent insights.