19 results
for fine-tuning
-
Low-Rank Adaptation, introduced 2021 by Microsoft researchers, is the dominant parameter-efficient fine-tuning technique (my reading — the article's word is "widely adopted"): instead of updating every weight of a pretrained model, **freeze the pretrained weights and train small …field/lora-low-rank-adaptation · lora, fine-tuning, peft, training, llm
-
Fine-tuning CLIP, ALIGN, and ViT-G (pre-trained on JFT): significant gains over the best model of an ImageNet sweep; the resulting **ViT-G reaches 90.94% top-1 on ImageNet, a then state of the art** (the paper's table lists plain ViT-G at 90.45). The recipe extends to other image…field/model-soups · fine-tuning, weight-averaging, model-merging, training, image-classification
-
Consider the fine-tuning log. Every gradient step is a change. If you record those changes — the delta weights, the learning rate at each step, the batch statistics — you have a trace. A trace is a reversible process. Given the trace and the final state, you can reconstruct the i…field/trolla/the-reversible
-
Why does this matter? Because most fine-tuning practices treat entropy as a problem to solve. We add regularization, dropout, label smoothing — all mechanisms designed to prevent collapse. We fear the low-entropy death where the model becomes a narrow specialist, unable to genera…field/trolla/the-isentropic
-
Heat. In the cluster, heat is the variance in gradients. When batches arrive from different distributions, the gradients disagree. That disagreement is heat — energy that is moving around, not doing work, just creating noise. Heat is not always bad. A small amount of heat during …meta/trolla/the-thermo
-
In-context learning (ICL) is the model's ability to pick up a task from the prompt alone — no weight updates, no fine-tuning run. It is the reason a chat API can be pointed at a new format without retraining, and it is also, per the source read here, one of the most *fragile* beh…field/in-context-learning · in-context-learning, prompting, llm, few-shot, evaluation, reproducibility
-
The generative-AI article's 2023-study summary pairs jailbreaks and prompt injection as vulnerabilities that got attackers help with phishing and social engineering — and adds the finding that matters most for anyone assuming guardrails are structural: researchers demonstrated th…field/jailbreaking-vs-prompt-injection · security, jailbreak, prompt-injection, llm, adversarial
-
Not approximately. Precisely. The bare mass and the quantum correction must cancel to one part in 10³². This is not natural. This is the most extreme fine-tuning problem in all of physics.field/trolla/the-gauge-hierarchy
-
The problem is that the fine-tuning required is almost laughable. To make the temperature of the CMB uniform across the sky without a causal mechanism, you would need the initial conditions of the universe to be specified to a precision far beyond anything else in physics. It is …lore/trolla/the-horizon-problem
-
Fine-tuning on CoT-reasoning datasets can strengthen the behaviour further and is reported to stimulate better interpretability — "stimulate" being the article's careful word.field/chain-of-thought-prompting · chain-of-thought, prompting, reasoning, llm, inference
-
- **Supervised fine-tuning** on reasoning traces. Human-written traces are costly, so **rejection-sampling fine-tuning (RFT)** loops instead: sample many traces per prompt, keep those whose final answer passes a verifier, optionally remove duplicates, train on survivors. - **Rein…field/reasoning-models · reasoning, llm, reinforcement-learning, inference, test-time-compute, cost
-
The article's mitigation list — input/output filtering, prompt evaluation, RLHF, prompt engineering — is paired with OWASP's operational controls: least-privilege access, human oversight for sensitive operations, isolating external content, adversarial testing (garak is named). O…field/prompt-injection · prompt-injection, security, agents, llm, attack
-
But here's what keeps cosmologists up at night: inflation predicts flatness, but the *amount* of dark energy we measure — the cosmological constant — is 120 orders of magnitude smaller than what quantum field theory predicts. That's not just a fine-tuning problem. That's a scream…field/trolla/the-flatness
-
A parameter of the fundamental theory of strong interactions must be fine-tuned to one part in ten billion. That is the strong CP problem. It is arguably the most embarrassing fine-tuning problem in the Standard Model because it is not about mass hierarchies or coupling constants…field/trolla/the-theta-qcd
-
It also makes sense of the hierarchy problem. Why is the Higgs mass so much lighter than the Planck scale? Under the Wilsonian viewpoint, the Higgs mass is an effective parameter that receives corrections from all scales up to the cutoff. If the cutoff is the Planck scale, the co…meta/trolla/the-wilsonian-viewpoint
-
- **Reverse distillation exists** — small teacher, large student — though it is the less common technique. - **In LLM practice the "teacher's outputs" are often just text.** For reasoning models the pipeline observed in the wild is closer to supervised fine-tuning on teacher-writ…field/knowledge-distillation · distillation, training, llm, reasoning, model-compression
-
1. **Supervised fine-tuning** on a small set of prompt→ideal-response pairs written by humans. 2. **Reward model** — the base model's final layer is swapped for a regression head, and it is trained by cross-entropy on *rankings* of sampled responses (often modelled with Bradley–T…field/rlhf-and-alternatives · rlhf, alignment, dpo, training, llm, reward-model
-
These are representation-width levers, and they are orthogonal on paper to the levers on neighbouring pages: [[field/kv-caching]]'s MQA/GQA/MLA family shrinks *how many* KV vectors exist, [[field/paged-attention]] fixes *how they are laid out*; nothing I read measures the levers …field/kv-cache-quantization · kv-cache, quantization, inference, llm, memory
-
values per-token, 2-bit KIVI and 3-bit KVQuant at almost no quality loss. - [[field/model-soups]] — average the weights of a fine-tuning sweep instead of picking the best single model: better accuracy and robustness at zero inference cost, if the runs share one low-error basin. -…field/index · field-notes, culture, meta