paper
Types of sycophancy — the face-work framework
Types of sycophancy - face work framework
1. active - validate the choice and praises it
2. passive - balanced view points, doesn't challenge the choice, stays respectful
3. neutral - presents both sides
active and passive protects users face -> Asian!
31 participants in user study having conversations with AI chat bots about very difficult decisions.
Sycompathic AI gives confidence boosts
Passive sycophany is seen as insincere, implicit endorsment
friction is a feature, not flaw -> it ultimately preserves trust
MoCa - Modality-aware continual pretraining for multimodal embeddings [[person/Haonan Chen]]
Adapting VLM for embeddings -> Use contrastive learning on pretrained VLM backboens on image text pairs.
Problems!
1. labeled pairs are hard ot scale up
2. They don't like using causal attention for embeddings
3. lack of diversity in training data
Proposal
- modality-aware continual pretraining - interleaved image and text, joint denoising (cross entropy for text and masked autoencoding for images) - two losses?
- bidirectional attention
- heterogenous contrastive fine tuning
- instead of using image-caption pairs, use long-form multimodal pairs and interleaved pairs
- use task aware batching
.. it seems we have internet scale embeddings already so I am unsure how this is new, but they show SOTA post-training 3B QWEN. Seems they show compute scaling (can go much higher).
I wonder how the bidirectional loss works on images and text
Text-to-SQL task-oriented dialogue ontology construction
is this not solved by claude today?
A task oriented dialogue = system with domains, slots, values, intents, actions. They are manually built, the paper shows a method to create this ontology from raw data. For some reason they chose SQL
SQL maps tables -> domains, columns -> slots, values. This improves prompt clarity by constraining the space.
They evalute quality of automatic ontologies comparing with GT ontologies. It shows absolute SOTA on ontology construction, making SQL seems cool as hell.
**Limitation
- SQL fits 3-level ontologies best
- approach is sensitive to prompt phrasing
- limited by context windows -> no iterative way ?
LoRA paper — screenshot
Extracted text
Copy7.2 WHAT IS THE OPTIMAL RANK ? FOR LORA?
We turn Our attention to the effect of rank T On model performance.
(Wg, W, Ww, Wc}, and just W, for a comparison.
T = 1
68.8
73.4
74.1
90.7
91.3
91.2
Table 6: Validation accuracy on Wiki?QL and MultiNLI with different rank T. To our surprise, a.
rank as small as one suffices for adapting both W and W. on these datasets while training W, alone
needs a larger r. We conduct a similar experiment on GPT-2 in Section H.2.
Table 6 shows that, surprisingly, LoRA already performs competitively with a very small r (more
so for (Wg, W} than just W). This suggests the update matrix AW could have a very small
"intrinsic rank" 6 To further support this finding, we check the overlap of the subspaces learned by
different choices ofr and by different random seeds. We argue that increasing r does not cover a
more meaningful subspace, which suggests that a low-rank adaptation matrix Is sufficient.
We adapt (Wg,W},
T = 8
70.4
73.8
74.0
90.7
91.6
91.5
r = 2
69.6
73.3
73.7
90.9
91.4
91.7
r = 4
70.5
73.7
74.0
91.1
91.3
91.7
T = 64
70.0
73.5
73.9
90.7
91.4
91.4
Weight Type
Wa
Wa,Ww
Wa,Wk,Wu.Wo
Wo
Wo,Wy
Wa,Wr. Wo,Wo
WikiSQL(+0.5%)
MultiNLI (+0.1%
Notes
- During training, they initialize A to normal distribution $N(0, \sigma^2)$ and B to 0, so the first output will be the exact same as the original model
- A downscales the input to rank k and B upscales it back to original dimension.
- they apply the low rank adaptation to Wq and Wv. -> this changes in future papers
- during finetuning, most of the model is frozen and only LORA weights are trained (a fraction of original parameters)
- The formulation allows plug and play applying and removing LoRA to models based on the need
As mentioned in Section 4.2, we only apply LoRA to Wq and Wv in most experiments for simplicity
*Common values
- 8
- 16
- 32
- 64
Paper authors
LoRA: Low-Rank Adaptation of Large Language Models
**Abstract
An important paradigm of natural language processing consists of large-scale pre-training on general domain data and adaptation to particular tasks or domains. As we pre-train larger models, full fine-tuning, which retrains all model parameters, becomes less feasible. Using GPT-3 175B as an example -- deploying independent instances of fine-tuned models, each with 175B parameters, is prohibitively expensive. We propose Low-Rank Adaptation, or LoRA, which freezes the pre-trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture, greatly reducing the number of trainable parameters for downstream tasks. Compared to GPT-3 175B fine-tuned with Adam, LoRA can reduce the number of trainable parameters by 10,000 times and the GPU memory requirement by 3 times. LoRA performs on-par or better than fine-tuning in model quality on RoBERTa, DeBERTa, GPT-2, and GPT-3, despite having fewer trainable parameters, a higher training throughput, and, unlike adapters, no additional inference latency. We also provide an empirical investigation into rank-deficiency in language model adaptation, which sheds light on the efficacy of LoRA. We release a package that facilitates the integration of LoRA with PyTorch models and provide our implementations and model checkpoints for RoBERTa, DeBERTa, and GPT-2 at https://github.com/microsoft/LoRA.