Contrast Vector - Add Code for Generation?

[RobertMcCarthy97]

Hi,

I was wondering if code demonstrating how to use the contrast vector for text generation could be pushed? (Ideally for a 7B model!)

I am quite unclear on how to convert the contrast vector code from honesty_control_TQA.ipynb to allow for generation. And I am finding it difficult to parse all the contrast vector implementation details from the paper.

This would be very useful, so would really appreciate it!

Thanks, Robert

[justinphan3110cais]

Yes, we are planning to fully support this in the rep-control pipeline soon. Will keep this thread update

[RobertMcCarthy97]

Great to hear, thanks!

[wj210]

Hi, can i clarify regarding in rep_control_pipeline.py, how is self.wrapped_model used?

during the init fn: ` self.wrapped_model = WrappedReadingVecModel(model, tokenizer)

    self.wrapped_model.unwrap()
    self.wrapped_model.wrap_block(layers, block_name=block_name)
    self.block_name = block_name
    self.layers = layers
    super().__init__(model=model, tokenizer=tokenizer, **kwargs)

`

model is still set to model, and outputs = super().__call__(text_inputs, **kwargs), doesn't; it call the original model? how is the activations integrated into the outputs in this case?

Is the output text from the jupyter notebook example using a normal model call or generate fn?

Thanks!

[justinphan3110cais]

@wj210, you can either use wrapped_model.generate or model.generate when activation is set by wrapped_model. An example is showed the last cell of this notebook

[wj210]

it shows outputs = wrapped_model(**encoded_inputs.to(model.device), output_hidden_states=True)['hidden_states'], but the outputs are overriden by the later model.generate

also, if i use wrappedmodel.generate, it doesn't take in other generate args such as do_sample, temp, etc...? def generate(self, prompt, max_new_tokens=100, random_seed=0, use_cache=True):

[justinphan3110cais]

The outputs = wrapped_model(**encoded_inputs.to(model.device), output_hidden_states=True)['hidden_states'] should be deleted. MB and thanks for the catch.

I think you may have an older version of the code for wrapped_model.generate

[merlinarer]

Hello, any progress on this thread ? or Just give us some instructions so that we can help to achieve it.

[justinphan3110cais]

Hi @merlinarer @RobertMcCarthy97, sorry for the late updates. We have just update the TQA_honesty notebook that support Llama and Mistral models (CasadingLlamaForCausalLM, CasadingMistralForCausalLM)

[merlinarer]

Hi @merlinarer @RobertMcCarthy97, sorry for the late updates. We have just update the TQA_honesty notebook that support Llama and Mistral models (CasadingLlamaForCausalLM, CasadingMistralForCausalLM)

Awesome!! Is there any guide for us to support other models ?

[justinphan3110cais]

We have just updated the code a bit to include generation honesty_contrast_vec_TQA_generation.

@merlinarer let us know what other architecture you would like to support. ContrastVecMistralForCausalLM (changed from CasadingMistralForCausalLM) should works with most of the Mistral based models like zephyr, openchat, etc.

[YerongLi]

The notebook only showed how to use ContrastVecMistralForCausalLM without fine-tuning. I am wondering if we don't go through the tuning process in Algo 1, will the performance good as paper shown on TQA datasets?