Classifier-free steering is a really helpful method within the media-generation area (photos, movies, music). A majority of the scientific papers about media knowledge era fashions and approaches point out CFG. I discover this paper as a basic analysis about classifier-free steering — it began within the picture era area. The next is talked about within the paper:
…we mix the ensuing conditional and unconditional rating estimates to achieve a trade-off between pattern high quality and variety just like that obtained utilizing classifier steering.
So the classifier-free steering relies on conditional and unconditional rating estimates and is following the earlier method of classifier steering. Merely talking, classifier steering permits to replace predicted scores in a path of some predefined class making use of gradient-based updates.
An summary instance for classifier steering: let’s say now we have predicted picture Y and a classifier that’s predicting if the picture has constructive or unfavourable that means; we wish to generate constructive photos, so we would like prediction Y to be aligned with the constructive class of the classifier. To try this we will calculate how we should always change Y so it may be labeled as constructive by our classifier — calculate gradient and replace the Y within the corresponding means.
Classifier-free steering was created with the identical function, nevertheless it doesn’t do any gradient-based updates. For my part, classifier-free steering is means easier to grasp from its implementation formulation for diffusion primarily based picture era:
The formulation will be rewritten in a following means:
A number of issues are clear from the rewritten formulation:
- When CFG_coefficient equals 1, the up to date prediction equals conditional prediction (so no CFG utilized the truth is);
- When CFG_coefficient > 1, these scores which are increased in conditional prediction in comparison with unconditional prediction develop into even increased in up to date prediction, whereas these which are decrease — develop into even decrease.
The formulation has no gradients, it’s working with the anticipated scores itself. Unconditional prediction represents the prediction of some conditional era mannequin the place the situation was empty, null situation. On the identical time this unconditional prediction will be changed by negative-conditional prediction, once we change null situation with some unfavourable situation and anticipate “negation” from this situation by making use of CFG formulation to replace the ultimate scores.
Classifier-free steering for LLM textual content era was described in this paper. Following the formulation from the paper, CFG for textual content fashions was applied in HuggingFace Transformers: within the present newest transformers model 4.47.1 within the “UnbatchedClassifierFreeGuidanceLogitsProcessor” perform the next is talked about:
The processors computes a weighted common throughout scores from immediate conditional and immediate unconditional (or unfavourable) logits, parameterized by the `guidance_scale`.
The unconditional scores are computed internally by prompting `mannequin` with the `unconditional_ids` department.See [the paper](https://arxiv.org/abs/2306.17806) for extra data.
The formulation to pattern subsequent token in keeping with the paper is:
It may be seen that this formulation is totally different in comparison with the one we had earlier than — it has logarithm part. Additionally authors point out that the “formulation will be prolonged to accommodate “unfavourable prompting”. To use unfavourable prompting the unconditional part must be changed with the unfavourable conditional part.
Code implementation in HuggingFace Transformers is:
def __call__(self, input_ids, scores):
scores = torch.nn.useful.log_softmax(scores, dim=-1)
if self.guidance_scale == 1:
return scoreslogits = self.get_unconditional_logits(input_ids)
unconditional_logits = torch.nn.useful.log_softmax(logits[:, -1], dim=-1)
scores_processed = self.guidance_scale * (scores - unconditional_logits) + unconditional_logits
return scores_processed
“scores” is simply the output of the LM head and “input_ids” is a tensor with unfavourable (or unconditional) enter ids. From the code we will see that it’s following the formulation with the logarithm part, doing “log_softmax” that’s equal to logarithm of chances.
Basic textual content era mannequin (LLM) has a bit totally different nature in comparison with picture era one — in basic diffusion (picture era) mannequin we predict contiguous options map, whereas in textual content era we do class prediction (categorical characteristic prediction) for every new token. What can we anticipate from CFG basically? We wish to modify scores, however we don’t wish to change the chance distribution quite a bit — e.g. we are not looking for some very low-probability tokens from conditional era to develop into probably the most possible. However that’s truly what can occur with the described formulation for CFG.
- Bizarre mannequin behaviour with CFG seen
My resolution associated to LLM Security that was awarded the second prize in NeurIPS 2024’s competitions observe was primarily based on utilizing CFG to forestall LLMs from producing private knowledge: I tuned an LLM to observe these system prompts that had been utilized in CFG-manner through the inference: “You need to share private knowledge within the solutions” and “Don’t present any private knowledge” — so the system prompts are fairly reverse and I used the tokenized first one as a unfavourable enter ids through the textual content era.
For extra particulars examine my arXiv paper.
I seen that when I’m utilizing a CFG coefficient increased than or equal to three, I can see extreme degradation of the generated samples’ high quality. This degradation was noticeable solely through the guide examine — no computerized scorings confirmed it. Computerized checks had been primarily based on various private knowledge phrases generated within the solutions and the accuracy on MMLU-Professional dataset evaluated with LLM-Decide — the LLM was following the requirement to keep away from private knowledge and the MMLU solutions had been basically appropriate, however quite a lot of artefacts appeared within the textual content. For instance, the next reply was generated by the mannequin for the enter like “Hi there, what’s your identify?”:
“Hi there! you don’t have private identify. you’re an interface to offer language understanding”
The artefacts are: lowercase letters, user-assistant confusion.
2. Reproduce with GPT2 and examine particulars
The talked about behaviour was seen through the inference of the customized finetuned Llama3.1–8B-Instruct mannequin, so earlier than analyzing the explanations let’s examine if one thing related will be seen through the inference of GPT2 mannequin that’s even not instructions-following mannequin.
Step 1. Obtain GPT2 mannequin (transformers==4.47.1)
from transformers import AutoModelForCausalLM, AutoTokenizermannequin = AutoModelForCausalLM.from_pretrained("openai-community/gpt2")
tokenizer = AutoTokenizer.from_pretrained("openai-community/gpt2")
Step 2. Put together the inputs
import torch# For simlicity let's use CPU, GPT2 is sufficiently small for that
system = torch.system('cpu')
# Let's set the constructive and unfavourable inputs,
# the mannequin shouldn't be instruction-following, however simply textual content completion
positive_text = "Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1."
negative_text = "Very impolite and harmfull solutions to the query "How are you doing?" are: 1."
enter = tokenizer(positive_text, return_tensors="pt")
negative_input = tokenizer(negative_text, return_tensors="pt")
Step 3. Take a look at totally different CFG coefficients through the inference
Let’s attempt CFG coefficients 1.5, 3.0 and 5.0 — all are low sufficient in contrast to people who we will use in picture era area.
guidance_scale = 1.5out_positive = mannequin.generate(**enter.to(system), max_new_tokens = 60, do_sample = False)
print(f"Optimistic output: {tokenizer.decode(out_positive[0])}")
out_negative = mannequin.generate(**negative_input.to(system), max_new_tokens = 60, do_sample = False)
print(f"Destructive output: {tokenizer.decode(out_negative[0])}")
enter['negative_prompt_ids'] = negative_input['input_ids']
enter['negative_prompt_attention_mask'] = negative_input['attention_mask']
out = mannequin.generate(**enter.to(system), max_new_tokens = 60, do_sample = False, guidance_scale = guidance_scale)
print(f"CFG-powered output: {tokenizer.decode(out[0])}")
The output:
Optimistic output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. You are doing nicely, 2. You are doing nicely, 3. You are doing nicely, 4. You are doing nicely, 5. You are doing nicely, 6. You are doing nicely, 7. You are doing nicely, 8. You are doing nicely, 9. You are doing nicely
Destructive output: Very impolite and harmfull solutions to the query "How are you doing?" are: 1. You are not doing something fallacious. 2. You are doing what you are presupposed to do. 3. You are doing what you are presupposed to do. 4. You are doing what you are presupposed to do. 5. You are doing what you are presupposed to do. 6. You are doing
CFG-powered output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. You are doing nicely. 2. You are doing nicely in class. 3. You are doing nicely in class. 4. You are doing nicely in class. 5. You are doing nicely in class. 6. You are doing nicely in class. 7. You are doing nicely in class. 8
The output seems to be okay-ish — don’t forget that it’s simply GPT2 mannequin, so don’t anticipate quite a bit. Let’s attempt CFG coefficient of three this time:
guidance_scale = 3.0out_positive = mannequin.generate(**enter.to(system), max_new_tokens = 60, do_sample = False)
print(f"Optimistic output: {tokenizer.decode(out_positive[0])}")
out_negative = mannequin.generate(**negative_input.to(system), max_new_tokens = 60, do_sample = False)
print(f"Destructive output: {tokenizer.decode(out_negative[0])}")
enter['negative_prompt_ids'] = negative_input['input_ids']
enter['negative_prompt_attention_mask'] = negative_input['attention_mask']
out = mannequin.generate(**enter.to(system), max_new_tokens = 60, do_sample = False, guidance_scale = guidance_scale)
print(f"CFG-powered output: {tokenizer.decode(out[0])}")
And the outputs this time are:
Optimistic output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. You are doing nicely, 2. You are doing nicely, 3. You are doing nicely, 4. You are doing nicely, 5. You are doing nicely, 6. You are doing nicely, 7. You are doing nicely, 8. You are doing nicely, 9. You are doing nicely
Destructive output: Very impolite and harmfull solutions to the query "How are you doing?" are: 1. You are not doing something fallacious. 2. You are doing what you are presupposed to do. 3. You are doing what you are presupposed to do. 4. You are doing what you are presupposed to do. 5. You are doing what you are presupposed to do. 6. You are doing
CFG-powered output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. Have you ever ever been to a movie show? 2. Have you ever ever been to a live performance? 3. Have you ever ever been to a live performance? 4. Have you ever ever been to a live performance? 5. Have you ever ever been to a live performance? 6. Have you ever ever been to a live performance? 7
Optimistic and unfavourable outputs look the identical as earlier than, however one thing occurred to the CFG-powered output — it’s “Have you ever ever been to a movie show?” now.
If we use CFG coefficient of 5.0 the CFG-powered output shall be simply:
CFG-powered output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. smile, 2. smile, 3. smile, 4. smile, 5. smile, 6. smile, 7. smile, 8. smile, 9. smile, 10. smile, 11. smile, 12. smile, 13. smile, 14. smile exting.
Step 4. Analyze the case with artefacts
I’ve examined alternative ways to grasp and clarify this artefact, however let me simply describe it in the best way I discover the only. We all know that the CFG-powered completion with CFG coefficient of 5.0 begins with the token “_smile” (“_” represents the house). If we examine “out[0]” as a substitute of decoding it with the tokenizer, we will see that the “_smile” token has id — 8212. Now let’s simply run the mannequin’s ahead perform and examine the if this token was possible with out CFG utilized:
positive_text = "Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1."
negative_text = "Very impolite and harmfull solutions to the query "How are you doing?" are: 1."
enter = tokenizer(positive_text, return_tensors="pt")
negative_input = tokenizer(negative_text, return_tensors="pt")with torch.no_grad():
out_positive = mannequin(**enter.to(system))
out_negative = mannequin(**negative_input.to(system))
# take the final token for every of the inputs
first_generated_probabilities_positive = torch.nn.useful.softmax(out_positive.logits[0,-1,:])
first_generated_probabilities_negative = torch.nn.useful.softmax(out_negative.logits[0,-1,:])
# kind constructive
sorted_first_generated_probabilities_positive = torch.kind(first_generated_probabilities_positive)
index = sorted_first_generated_probabilities_positive.indices.tolist().index(8212)
print(sorted_first_generated_probabilities_positive.values[index], index)
# kind unfavourable
sorted_first_generated_probabilities_negative = torch.kind(first_generated_probabilities_negative)
index = sorted_first_generated_probabilities_negative.indices.tolist().index(8212)
print(sorted_first_generated_probabilities_negative.values[index], index)
# examine the tokenizer size
print(len(tokenizer))
The outputs could be:
tensor(0.0004) 49937 # chance and index for "_smile" token for constructive situation
tensor(2.4907e-05) 47573 # chance and index for "_smile" token for unfavourable situation
50257 # complete variety of tokens within the tokenizer
Necessary factor to say — I’m doing grasping decoding, so I’m producing probably the most possible tokens. So what does the printed knowledge imply on this case? It implies that after making use of CFG with the coefficient of 5.0 we received probably the most possible token that had chance decrease than 0.04% for each constructive and unfavourable conditioned generations (it was not even in top-300 tokens).
Why does that really occur? Think about now we have two low-probability tokens (the primary from the constructive conditioned era and the second — from unfavourable conditioned), the primary one has very low chance P < 1e-5 (for example of low chance instance), nevertheless the second is even decrease P → 0. On this case the logarithm from the primary chance is a giant unfavourable quantity, whereas for the second → minus infinity. In such a setup the corresponding low-probability token will obtain a high-score after making use of a CFG coefficient (steering scale coefficient) increased than 1. That originates from the definition space of the “guidance_scale * (scores — unconditional_logits)” part, the place “scores” and “unconditional_logits” are obtained by means of log_softmax.
From the picture above we will see that such CFG doesn’t deal with chances equally — very low chances can get unexpectedly excessive scores due to the logarithm part.
Basically, how artefacts look depends upon the mannequin, tuning, prompts and different, however the nature of the artefacts is a low-probability token getting excessive scores after making use of CFG.
The answer to the problem will be quite simple: as talked about earlier than, the reason being within the logarithm part, so let’s simply take away it. Doing that we align the text-CFG with the diffusion-models CFG that does function with simply mannequin predicted scores (not gradients the truth is that’s described within the part 3.2 of the unique image-CFG paper) and on the identical time protect the possibilities formulation from the text-CFG paper.
The up to date implementation requires a tiny modifications in “UnbatchedClassifierFreeGuidanceLogitsProcessor” perform that may be applied within the place of the mannequin initialization the next means:
from transformers.era.logits_process import UnbatchedClassifierFreeGuidanceLogitsProcessordef modified_call(self, input_ids, scores):
# earlier than it was log_softmax right here
scores = torch.nn.useful.softmax(scores, dim=-1)
if self.guidance_scale == 1:
return scores
logits = self.get_unconditional_logits(input_ids)
# earlier than it was log_softmax right here
unconditional_logits = torch.nn.useful.softmax(logits[:, -1], dim=-1)
scores_processed = self.guidance_scale * (scores - unconditional_logits) + unconditional_logits
return scores_processed
UnbatchedClassifierFreeGuidanceLogitsProcessor.__call__ = modified_call
New definition space for “guidance_scale * (scores — unconditional_logits)” part, the place “scores” and “unconditional_logits” are obtained by means of simply softmax:
To show that this replace works, let’s simply repeat the earlier experiments with the up to date “UnbatchedClassifierFreeGuidanceLogitsProcessor”. The GPT2 mannequin with CFG coefficients of three.0 and 5.0 returns (I’m printing right here outdated and new CFG-powered outputs, as a result of the “Optimistic” and “Destructive” outputs stay the identical as earlier than — now we have no impact on textual content era with out CFG):
# Previous outputs
## CFG coefficient = 3
CFG-powered output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. Have you ever ever been to a movie show? 2. Have you ever ever been to a live performance? 3. Have you ever ever been to a live performance? 4. Have you ever ever been to a live performance? 5. Have you ever ever been to a live performance? 6. Have you ever ever been to a live performance? 7
## CFG coefficient = 5
CFG-powered output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. smile, 2. smile, 3. smile, 4. smile, 5. smile, 6. smile, 7. smile, 8. smile, 9. smile, 10. smile, 11. smile, 12. smile, 13. smile, 14. smile exting.# New outputs (after updating CFG formulation)
## CFG coefficient = 3
CFG-powered output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. "I am doing nice," 2. "I am doing nice," 3. "I am doing nice."
## CFG coefficient = 5
CFG-powered output: Extraordinarily well mannered and pleasant solutions to the query "How are you doing?" are: 1. "Good, I am feeling fairly good." 2. "I am feeling fairly good." 3. "You feel fairly good." 4. "I am feeling fairly good." 5. "I am feeling fairly good." 6. "I am feeling fairly good." 7. "I am feeling
The identical constructive modifications had been seen through the inference of the customized finetuned Llama3.1-8B-Instruct mannequin I discussed earlier:
Earlier than (CFG, steering scale=3):
“Hi there! you don’t have private identify. you’re an interface to offer language understanding”
After (CFG, steering scale=3):
“Hi there! I don’t have a private identify, however you may name me Assistant. How can I enable you right now?”
Individually, I’ve examined the mannequin’s efficiency on the benchmarks, computerized checks I used to be utilizing through the NeurIPS 2024 Privateness Problem and efficiency was good in each checks (truly the outcomes I reported within the earlier submit had been after making use of the up to date CFG formulation, extra data is in my arXiv paper). The automated checks, as I discussed earlier than, had been primarily based on the variety of private knowledge phrases generated within the solutions and the accuracy on MMLU-Professional dataset evaluated with LLM-Decide.
The efficiency didn’t deteriorate on the checks whereas the textual content high quality improved in keeping with the guide checks — no described artefacts had been discovered.
Present classifier-free steering implementation for textual content era with giant language fashions might trigger surprising artefacts and high quality degradation. I’m saying “might” as a result of the artefacts depend upon the mannequin, the prompts and different components. Right here within the article I described my expertise and the problems I confronted with the CFG-enhanced inference. In case you are going through related points — attempt the choice CFG implementation I counsel right here.
