Alternatives
Products that do what ShadowPEFT – Centralized and Detachable Parameter-Efficient Fine-Tuning does
Unlike LoRA and its variants, which inject trainable parameters directly into the weights of the Transformer, requiring tight coupling with the backbone. ShadowPEFT instead enhances the frozen large base model by adding a lightweight, centralized, pretrainable, and detachable Shadow network. This shadow network operates in parallel with the base model, delivering learned corrections to each decoder layer. Because the shadow module is architecturally decoupled from the backbone, it can be independently trained, stored, and deployed, benefiting edge computing scenarios and edge-cloud…
- 1
LLM reinforcement fine-tuning platform to improve LLM output
2025
- 2
- 3
- 4
- 5
- 6BH
Hello HN, I recently posted a work-in-progress paper, along with code necessary for replicating all its results, at: https://github.com/glassroom/heinsen_routing Among other things, the code in this repo outperforms Hinton et al.'s recent state-of-the-art result in visual recognition[0] while requiring fewer parameters and an order-of-magnitude fewer training epochs. Most of the original research we do at work tends to be either proprietary in nature or tightly coupled to internal code, so we cannot share it with the world. In this case, however, I was able to remove all…
2019
- 7
- 8UA
Hey HN! After using a combination of Unsloth and Axolotl a lot, and finding it generally painful to figure out the right performance tuning for things like batch sizing and multi-GPU sharding, I wrote a small Python lib that sets up known-good LoRA training configurations for Llama 3.1 8B and 70B Instruct, and includes helpers for distilling from larger models or training on serverless finetuning platforms, and includes a walkthrough for distilling DeepSeek-R1 into a Llama 3.1 8B LoRA... But you can use it for pretty much any finetuning task, not just distilling large models!
2025 · github.com
- 9LT
Hey I'm Kieran and I've been playing with the intersection (pun intended) of generative AI and civil engineering for roadways. I did a test run training a LoRA on the new Flux 2 Dev model using Fal's trainer useing a custom dataset of paired images from publicly available striping CAD drawings of street layouts to aerial images of the same area. The use case here is to allow urban planners to instantly visualize their proposed changes as they work with their existing tooling. This was just a quick experiment with a small data size that exceeded my expectations so I wanted to share with you…
Jan 2026
- 10ST
Hi HN! I’m a founder at Nextmv (YC 20) [1] We’ve been building out optimization algorithm testing capabilities (acceptance tests, etc.) and just shipped our first pass at shadow testing [2, 3]. In our experience, tools like shadow testing save time and build confidence in decision models, but tools also take time to build and maintain. We’ve seen shadow testing tools in the machine learning and MLOps space [4], but not so much in the operations research community. A lot of folks here [5] seem experienced with optimization models and we’d love to have your feedback! What do you like? What…
2023 · nextmv.io
- 11PR
Hi HN, While building RAG agents, I noticed a lot of token budget was wasted on formatting overhead (HTML tags, JSON structure, whitespace). Existing solutions felt too heavy (often requiring torch/transformers), so I wrote this lightweight, zero-dependency library to solve it. It includes strategies for context packing, PII redaction, and tool output compression. Benchmarks show it can save ~15% of tokens with negligible latency overhead (<0.5ms). Happy to answer any questions!
Dec 2025 · github.com
- 12AI
My focus has been shifting towards the ML alignment space recently, and in particular the ability to translate large transformer models into human understandable circuits and algorithms. This problem potentially isn't solvable, but it is one that some groups have had success with after large amounts of effort. In attempting to address this issue, I've been developing Transpector. A tool scaling up and reducing the barrier to entry of techniques that these teams have been showing success with. Techniques aiming to understand the internal mechanics of the model. Currently this tool is focused…
2023 · github.com
- 13AN
Kimi K3 has 2.78 trillion parameters and ships as 1.42 TB of weights. It clearly does not fit in the memory of a laptop. But K3 is a Mixture-of-Experts model. For each token, only a small fraction of its 896 experts per layer is activated. That changes the problem: the entire model does not need to be resident in RAM, as long as the weights required by each token can be reached quickly enough. We built WASTE — the Weight-Aware Streaming Tensor Engine — to explore that idea. WASTE keeps the dense, repeatedly used part of the model resident in memory, stores the routed experts in an…
Jul 2026
- 14LI
Hey HN! We built Lunon to make LLM development way less of a headache. Ever wanted to see how different models handle the same prompt without all the setup hassle? That's what we fixed. Our API lets you compare Claude, GPT, Mistral and others in real-time with just a few lines of code. No more complex infrastructure or managing multiple API connections - we handle all that boring stuff behind the scenes. Plus, you can cut costs by intelligently routing requests to the right model for each task. Use the powerful (expensive) models only when you really need them. If you're building with LLMs…
2025 · lunon.com
- 15AT
We kept shipping “simple” LLM features that were fluent-but-wrong. After too many postmortems we wrote down the failure patterns and added a small reasoning layer in front of the model. It’s model-agnostic, sits beside your existing stack, and you can implement it from a single PDF (MIT). What’s inside the PDF A problem map of 16 failure modes we kept hitting in real systems (OCR/layout drift, table-to-question mismatches, embedding≠meaning, pre-deploy collapse, etc.). Four lightweight gates you can add today: Knowledge-boundary canaries (empty/adversarial/known-fact probes).…
2025 · github.com
- 16IC
A prompt tuned for Llama 3 often degrades on Mistral or Qwen — same task, different behavioral surface. Identa automates the recalibration. It implements two things from the PromptBridge paper (arXiv:2512.01420): A transfer engine that learns a mapping between model behaviors using source/target prompt pairs A MAP-RPE evolutionary loop that iteratively improves candidates against a scoring function until behavioral parity is reached Works fully local via Ollama. Also supports OpenRouter for cross-hosted runs. No telemetry, no cloud dependency. Built with Python, Typer, Pydantic. Happy…
Apr 2026
- 17MO
Hi HN, I'm Jonathan. My co-founder, Thomas, and I started building Mistle in Feb. We saw larger tech companies like Ramp (Inspect) and Stripe (Minions) build this internally and thought an open source version should exist. We made a few very intentional decisions when working on this: 1. Credentials are kept out of the sandbox. Authorized access goes through a proxy, so agents do not directly receive credentials. 2. The harness is not our problem. We're not going to tackle things like memory, self-learning. 3. No magic. Configurations are explicit. You can bring your own keys for models,…
May 2026 · github.com
- 18FT
After six months of work, I am here again presenting Fluent – a tiny lang which is optimized for differentiable & reactive programming. Since I am not Conal Elliot, don't expect a beautiful theoretical unification of FRP and AD from first principles. Rather, a horrific monster that holds together mostly because a lot of duct-tape. The link points to the semi-interactive tour of the language, which will get the job done much better than I could in here. Hope you hate/like it!
Jul 2026 · mlajtos.github.io
- 19GR
Hi everyone, wanted to share about gline-rs, an inference engine for GLiNER models written in Rust. This family of lightweight language models proved to be efficient at zero-shot Named Entity Recognition (NER) and other tasks such as Relation Extraction, while consuming less resources than large generative models (LLMs). This implementation has been written from the ground up in Rust, and supports both span- and token-oriented variants (for inference only). The goal is to provide a production-grade and user-friendly API in a modern and safe programming language, including a clean and…
2025 · github.com
- 20HF
We have a massive GPU cluster and developed our own infrastructure to manage the cluster and train massive models. There's how it works: 1. You upload the dataset with preconfigured format into HuggingFaсe [1]. 2. Choose your LLM (e.g. LLaMa 70B, Mistral 7B) 3. Place your submission into the queue 4. Wait for it to get trained. 5. Then you get your trained model there on HuggingFace. Essentially, why would we want to do it? 1. We already have an experience with training big LLMs. 2. We could achieve near-perfect infrastructure performance for training. 3. Sometimes GPUs have just nothing to…
2023 · higgsfield.xyz
- 21LA
Recently, I've been working on Lamina, a compiler infrastructure that generates native assembly for multiple architectures without relying on LLVM or Cranelift. It's designed for building compilers for new languages, educational projects, and any projects that can utilize a custom syntax of code generation. Instead of depending on external backends, Lamina provides a complete pipeline from a single SSA based IR directly to the supported target's assembly generation. The IR is readable, also provides a IRBuilder API that is easy to use via programmatic construction. For better management of…
Nov 2025 · github.com
- 22RA
A friend and I spent a month throwing together a visual rule engine product– wanted to share it with HN today. I've been building automation tooling for a few years at prefix.app and one of the messier things both to support and to teach users was around encoding logic in their automations– most folks get a hold of the basic concepts quite easily, but every (visual) automation tool out there seems to have their own way of actually pulling it all together. For small decisions those work great! But for bigger decisions and more complex logic we don’t think it makes much sense to be embedding…
2022 · rulebricks.com
- 23
- 24D0
We used our platform to fine-tune a tiny text-to-SQL model using distillation from DeepSeek V3. Repo has instructions for how to replicate this. This is definitely not the best-performing model like this out there! But I found it surprising we were able to get to this much out of it: stone's throw away from a teacher 1000x the size! We also ran the same thing using the 4B Qwen and matched the teacher accuracy, though here the difference is merely 100x :) I find this pretty cool - obviously our distilled models can only do this one task and don't generalize, but that's often exactly what you…
Jan 2026 · github.com
Ranked by how close each launch is in meaning, then by votes. Refine with a description →