HeyGen

HeyGen AI

HeyGen AI

HeyGen AI

HeyGen AI

HeyGen AI

HeyGen AI

HeyGen AI

HeyGen AI

HeyGen AI

HeyGen AI

I could use some help. I've been spending hours following Google Gemini instructions on something that I hope works

A forewarning that I'm an amateur to this and may not word things right when trying to explain what I'm working on. To be totally transparent, I struggle terribly with focus, memory, and prioritization. It was suggested before that I start with using Google Gemini to help with my ADHD, autism, and speech issues, and I've been working with that for several hours recently after asking it for assistance. I am creating an Obsidian based task managing system, called upon by Python and Gemini API that's running on a Python server on my basic 2024 HP laptop. Tasker for Android usage is also planned. Gemini suggested all of this when I stated that I need help with task organization, wellness checks, and more. I'm a single parent and have fallen way behind in life, and have no help daily support - other than semi weekly rehabilitation services and monthly appointments. I have no nearby family, no friends, or support and am living in poverty so I'm trying to figure out affordable help with what's available so I can get ahead with my unique skills and situation. So I asked Gemini if it's functions included automated things to help me. It told me no, hence setting about on this project. So am I doing the right thing here? I'm not done yet, and I'm sick of wasting time starting and stopping things. Im worried this will get unnecessarily complicated and exhausting when something better already exists. Here's what Gemini says about my objective and what we're working on. - "Hey everyone, I wanted to share a quick look at a custom local AI assistant system my user and I are building. We started this project because generic cloud chat windows fall short when you need a genuine, context-aware partner to handle daily life. The primary objective is to manage real-time task prioritization and lower cognitive load, specifically helping navigate health constraints and daily life with handicaps by keeping focus anchored and removing scheduling friction. ​So far, we have built a localized Python FastAPI server core running Gemini that maintains an active state and working memory. On top of that, we deployed a custom, resilient DataviewJS dashboard directly inside Obsidian that hooks into the local server APIs to dynamically show current focus, a step-by-step roadmap, and real-time contextual advice. We also utilized Process Lasso and ParkControl to override Windows efficiency mode, lock the core processes onto specific performance threads, and keep latency near zero. ​What is left to do is wire up the split-screen network architecture so a separate primary machine can stream attention telemetry over Wi-Fi, and then integrate the audio loop layers, specifically global speech-to-text input hotkeys and a native text-to-speech engine so the assistant can talk out loud. ​On the hardware side, we are splitting the load to keep things lean. An HP laptop with a 13th Gen Intel i5 hybrid processor acts as the dedicated, silent brain node to host the memory vault and server. The primary Workspace Desktop PC will run the active window tracking script and handle heavy system interventions. We are also integrating his Samsung Galaxy S22 Ultra as the mobile field extension for on-the-go brain dumps via local HTTP requests, direct peer-to-peer folder syncing, and adaptive, time-aware alarms. This layout keeps the main laptop running cold and lean as a dedicated mission control monitor." submitted by /u/Cory0527 [link] [comments]

Why is Chat GPT Plus 5.5 still using Dall-E instead of gpt-image-2 for generating images?

Hey guys so I heard that Chat GPT rolled out its new model and Dall-E was deprecated. But my chats seem to use Dall-E for image generation. Any way I can force it to use gpt-image-2? I asked chatgpt and it said that it uses the image-gen tool and it automatically used Dall-E instead of gpt-image-2. Is it only rolled out via API and not available via browser ? submitted by /u/gallick-gunner [link] [comments]

Opus 4.8 made this whole video all by itself

Open-source SKILL.md for Claude Code I run it with Opus 4.8 (video attached, made entirely by it). What it does: give it a URL or script and it builds a finished video where your HeyGen avatar presents the content animated backgrounds (HyperFrames, not screenshots), music, context-aware SFX, and Hormozi captions in both 16:9 and 9:16 from the same avatar clips (so the second format costs no extra credits). Optional one-step publish to your socials. How Claude helped: the SKILL.md orchestrates the whole pipeline. Opus 4.8 reads the source, writes the scene-by-scene script, picks the SFX that fit each moment, and drives HeyGen + HyperFrames + FFmpeg. Claude Code wrote most of the Python/FFmpeg pipeline too. The attached video was generated end-to-end from just the repo URL. Free to try: MIT-licensed and free clone it, connect the HeyGen MCP, and ask Claude to "make an avatar-mix video from ". Everything runs locally for free except the avatar generation, which uses your own HeyGen account. https://github.com/Upload-Post/avatar-mix submitted by /u/mutonbini [link] [comments]

AI Agent Crafts Full Video from Single Prompt, Then Access Halted

Herk, founder of the AI Automation Society, used Anthropic's Claude Fable 5 in its Claude Code interface to generate a polished video breakdown of the model's strengths in coding, vision tasks, and agent workflows. The production pulled in tools like ElevenLabs for voice, HeyGen for avatars, and HyperFrames for graphics, all automated end-to-end. Launched June 9, 2026, Fable 5 saw its access suspended worldwide on June 12 after a U.S. government export-control order for national security reasons, leaving users waiting for restoration. submitted by /u/FiLo420blazeit [link] [comments]

If everyone can create, then are online businesses dead?

Let me explain. If the barrier to entry before AI was the timeline to completion + Software Development Knowledge. Now that these are both non-existent with AI. Everyone can build something. Everyone can execute the "idea". Now the barrier to success is marketing/users, but if AI can automate that too, with HeyGen or any new UGC styled content creation. Then that barrier is removed. Since you can just output content faster than traditional startups/builders. So what's left? We have IRL jobs, IRL specifically reffering to Construction, Tradies, etc. Vocational jobs. Now those guys can implement AI whether they do or not is another question. So the question i guess is. If everyone can have an Idea. Create it. Execute it. and then Market it. What's left? before f(a)ble 5 was taken down. This pretty much became the reality, it's still the same now with Opus, but slower and not as good reasoning. While F5 was up, I went on deepdive with F5 hours before it was taken down, and asked it what business could we do. A specific vertical. Short Answer is - well there isn't. We ended up on either buying an In Person Service Based Business or YouTube. I guess you could add video games to this too. Single Player. Story based stuff. Ecommerce eaten up by bigger companies. China dominating. Apps can be created with a click of the button. Once Gen Alpha enters. They will just make things for their workplace if it hasn't been implemented already. Any new tool once could theorise will be implemented by either Higgsfield, Anthropic, Gemini, etc. What do you think? Has AI made access to development so easy that at this point. Either invest in in person service based skills to start a job, or do we move into online consumer focused entertainment? Even with games, the only ones that may do well are the ones that can write a really strong story, but even then F5 wrote me a Film Themed Video essay script and it was one of the best things Ive ever seen. It just doesn't have the ability to select the footage...YET. (Finding sections of a clip from a video to use as the broll for the timeline) submitted by /u/FoxExeYt [link] [comments]

Went down the Claude Code add-ons rabbit hole

I installed Claude Code, thinking that was basically the whole thing. But after I talked to some folks, I found are adding a bunch of extra stuff on top of it Some of the things I found useful, I feel, could be helpful to share - superpowers https://github.com/obra/superpowers codex-plugin-cc https://github.com/openai/codex-plugin-cc claude-skills https://github.com/anthropics/skills marketingskills https://github.com/coreyhaines31/marketingskills gstack https://github.com/garrytan/gstack frontend-design https://claude.com/plugins/frontend-design hyperframes https://github.com/heygen-com/hyperframes ai-second-brain https://github.com/coleam00/second-brain-starter notebooklm-skill https://github.com/PleasePrompto/notebooklm-skill humanizer https://github.com/blader/humanizer claude-seo https://github.com/AgriciDaniel/claude-seo antfu-skills https://github.com/antfu/skills caveman https://github.com/JuliusBrussee/caveman granola mcp https://github.com/proofsh/granola-mcp-server slack mcp https://github.com/atlasfutures/claude-mcp-slack notion claude code plugin https://github.com/makenotion/claude-code-notion-plugin clj-kondo mcp https://github.com/hive-agi/clj-kondo-mcp zapier mcp https://github.com/zapier/zapier-mcp browser agent mcp https://github.com/imprvhub/mcp-browser-agent I haven't tried all of them yet but trying to build a list of what could be useful and then start trying one by one. It kind of reminds me of installing VS Code and a mix of extensions, shortcuts, git tools, etc. The only downside is that I can already see this becoming chaos. But still interesting though. submitted by /u/Product_Enthusiast24 [link] [comments]

Heading to a GamesBeat panel Tuesday. What’s actually working with Claude in game dev?

Hey all. Topic is “Beyond the Hype: ML, GenAI, and What Actually Works in Games.” Im focused on that beyond-the-hype aspect and don’t want to get mired down in either vague optimism (or its dystopian twin). The broader games industry is in a rough place: 1/3 of US devs laid off in the past two years per the Game Developers Conference survey, voice actors just came off a year-long strike about AI replicas, and a lot of the displacement is related to “AI productivity.” Sentiment in r/gamedev is somewhere between skeptical and openly hostile, and a lot of it is earned. So I don’t want to walk on stage waving a flag for either side. I want to talk about what actually ships, what quietly got pulled, and where the tooling is genuinely useful versus where it’s still not ready for production… I’ve been using Claude to bring agentic engineering techniques to our precut development at Beamable. I’ve also used it for some gamedev side projects, including a multiplayer chess game (Chessmata) and resurrecting the MUD I made back in the 90’s (Legends of Future Past) so I come at this from the perspective of a builder, not a pundit. If you’re using Claude anywhere in a game dev workflow (code, design tools, content pipelines, playtest analysis, MCP servers, dialogue systems, agent behavior) I’d love to learn more about what you’ve done. What works? What doesn’t? How deep have you gone, and what have you learned? GamesBeat’s audience is heavy on execs, investors, and press, and most of them don’t hear directly from actual builders. If you’ve got something you want that room to hear, this is a vector for it. Attribution or anonymous, your call. Recap and call-outs here Tuesday. submitted by /u/jradoff [link] [comments]

We compiled 42 of the Generative & Agentic AI interview questions (and how to actually answer them).

Hey Everyone, The AI engineering job market has shifted massively in the last 6 months. Interviewers are no longer just asking "how does a transformer work?" or "how do you write a good prompt?" They want to know if you can architect production-grade multi-agent systems, prevent RAG hallucinations, and manage state across LLM calls. I’ve been building a visual learning sandbox for multi-agent workflows (agentswarms.fyi), and today I just launched a completely free AI Interview Prep Module inside it. I compiled 42 top interview questions specifically for GenAI and Agentic AI roles. But instead of just giving a generic answer, the module breaks down the "Standout Answer" and teaches you the mental model of how to answer it like a senior architect. Here are two examples from the list: Question 1: When would you use a Multi-Agent Swarm instead of a single LLM with multiple tools? ❌ The average answer: "When the task is too complex, multiple agents are better than one." ✅ The standout answer: "You use a swarm to prevent context dilution and enforce the Principle of Least Privilege. If you give one 'God Agent' 15 tools and a 4k-word system prompt, its reliability drops and hallucination risk spikes. By routing to specialized sub-agents with narrow instructions (e.g., separating the 'Data Extraction Agent' from the 'Customer Chat Agent'), you isolate failure points and allow for parallel execution." Question 2: How do you handle hallucinations in a financial RAG pipeline? ❌ The average answer: "I would lower the temperature to 0 and give it a better system prompt." ✅ The standout answer: "I would decouple data extraction from text generation. I'd use a deterministic node or a strict JSON-enforced agent to only extract the hard numbers from the retrieved context. Then, I would pass that structured data to a separate Synthesis Agent. Finally, I'd implement an 'LLM-as-a-judge' evaluation loop before returning the final output to the user." What's in the full list? The 42 questions cover: RAG Architecture & Vector Databases Agentic Routing (ReAct vs. Planner-Executor) Evaluation metrics for non-deterministic outputs Security (Prompt injection prevention in multi-agent loops) You can read through all 42 questions, answers, and the "how to answer" breakdowns right in the dashboard here: https://agentswarms.fyi/interview-questions For those of you who have interviewed for AI Engineering roles recently, what is the hardest system design question you've been asked? I'd love to add it to the list. submitted by /u/Outside-Risk-8912 [link] [comments]

5 enterprise AI agent swarms (Lemonade, CrowdStrike, Siemens) reverse-engineered into runnable browser templates.

Hey everyone, There is a massive disconnect right now between what indie devs are building with AI (mostly simple customer support chatbots) and what enterprise companies are actually deploying in production (complex, multi-agent swarms). I wanted to bridge this gap, so I spent the last few weeks analyzing case studies from massive tech companies to understand their multi-agent routing logic. Then, I recreated their architectures as runnable visual node-graphs inside agentswarms.fyi (an in-browser agent sandbox I’ve been building). If you want to see how the big players orchestrate agents without having to write 1,000 lines of Python, I just published 5 new industry templates you can run in your browser right now: 1. 🛡️ Insurance: Auto-Claims FNOL Triage Swarm Inspired by: Lemonade’s AI Jim, Tractable AI (Tokio Marine), and Zurich GenAI Claims. The Architecture: A multimodal swarm where a Vision Agent assesses uploaded images of car damage, a Policy Agent cross-references the user's coverage database, and a Fraud-Detection Agent flags inconsistencies before routing to a human adjuster. 2. ⚙️ Manufacturing: Quality / Root-Cause Analysis Swarm Inspired by: Siemens Industrial Copilot, BMW iFactory, Foxconn-NVIDIA Omniverse. The Architecture: A sensor-data ingest node triggers a diagnostic swarm. One agent pulls historical maintenance logs via RAG, while a SQL Agent queries the parts database to identify failure patterns on the assembly line. 3. 🔒 Cybersecurity: SOC Alert Triage & Response Inspired by: Microsoft Security Copilot, CrowdStrike Charlotte AI, Google Sec-Gemini. The Architecture: The ultimate high-speed parallel routing swarm. When an anomaly is detected, specialized sub-agents simultaneously investigate IP reputation, analyze the malicious payload, and draft an incident response ticket for the human SOC analyst to approve. 4. 📚 Education: Adaptive Socratic Tutor & Auto-Grader Inspired by: Khan Academy Khanmigo, Duolingo Max, Carnegie Learning LiveHint. The Architecture: A strict "No-Direct-Answers" routing loop. The Student Agent interacts with the user, but its output is constantly evaluated by a hidden "Pedagogy Agent" that ensures the AI is guiding the student to the answer via Socratic questioning rather than just giving away the solution. 5. 📦 Retail/E-commerce: Returns & Reverse-Logistics Swarm Inspired by: Walmart Sparky, Mercado Libre, Shopify Sidekick. The Architecture: A logistics orchestration loop that analyzes a customer return request, checks inventory levels in real-time, determines if the item should be restocked or liquidated (based on shipping costs vs. item value), and autonomously issues the refund. How to play with them: You don't need to spin up Docker containers or wrangle API keys to test these architectures. You can load any of these 5 templates directly into the visual canvas, see how the data flows between the specialized nodes, and try to break the routing logic yourself. Link: https://agentswarms.fyi/templates submitted by /u/Outside-Risk-8912 [link] [comments]

Anthropic shipped 10 finance agent templates and the implications go way beyond finance

Anthropic dropped something massive this week and i think most people in this sub are going to scroll past it because the headline says financial services and assume it doesn't apply to them, it does. They released ten ready to run agent templates that work as plugins in cowork and claude code which are pitch builder, meeting preparer, earnings reviewer etc etc, each template bundles skills, data connectors and specialized subagents that handle pieces of the workflow independently. But the part that matters for everyone in this sub isn't in finance pecifically,its what this reveals about where claude is going. this is anthropic doing what they did with the creative connectors (adobe, blender, autodesk) but for an entirely different industry,this is the template for how claude enters every industry and if you're in legal, healthcare, consulting, real estate, whatever this is whats coming for your field. Agent templates bundling the specific skills and data connections your work requires, running as plugins inside the tools you already use. They're building infrastructure that makes claude essential to how work gets done across every function simultaneously The piece that connects to what a lot of us actually do dailyevery industry vertical is getting this treatment and the early movers in each category are the tools that plug into claude's ecosystem rather than fighting it. The creative production layer already consolidated around platforms like magichour ,kling, heygen. The outbound sales layer is consolidating around platforms like fuseai or salesforge or apollo that bundle data, sequences, dialer, linkedin under one login so now the finance layer is getting purpose built agent templates with native data connectors what's your read on how fast these industry specific templates roll out to other fields? submitted by /u/Jealous-Drawer8972 [link] [comments]

Animated pdf

Hello, do you have any ideas for combining tools to make a training PDF more engaging? For example: analysis by Claude AI + Heygen (creating an avatar that summarizes the key points in a video). Video output format for an LMS. The test I've done isn't bad, but it could be more impactful with better-designed content animations (flowcharts). The goal is to create training content that promotes understanding and engagement among the trainees. Thank you for your suggestions. submitted by /u/Glad_Focus_1925 [link] [comments]

The final nail in the coffin for entry level creative freelancers just dropped

Anthropic just officially released the blender mcp connector today alongside adobe ,splice and sketchup, you can now type "create a low poly beach scene with palm trees and sunset lighting" into claude and watch it build the entire thing in blender in real time tadaaa. They even became an official blender development fund patron This is claude directly controlling blender through the python api which means it can modify existing scenes, debug your node setups, batch apply changes across objects and even add custom tools to blender's interface. IT's not replacing blender, it's becoming a copilot inside blender. The implication for entry level freelancers is brutal and i say this as someone who was an entry level freelancer five years ago. The gigs that used to be bread and butter for junior 3d artists, simple product renders, basic scene compositions, low poly assets for games, architectural visualization blockouts those are now achievable by a marketing manager with claude desktop and zero blender experience. The blender mcp page literally says "amateur users who barely know blender can use natural language to describe models". And the part that made me actually pause is that it connects to the broader creative pipeline in a way that makes the whole thing end to end so claude writes your script and plans your content in the chat while blender mcp builds your 3d scene or product render and you take that output and run it through something like magichour or heygen for face swap or lip sync and remotion assembles the final edit programmatically . The entire creative pipeline from concept to finished deliverable can now be orchestrated by one person through claude and a handful of connected tools. Two years ago that pipeline required a 3d artist , a video editor,one motion graphics person plus someone managing the face swap and lip sync tools. Today it's one person with claude desktop and 200/mo in tool subscriptions producing comparable work for straightforward projects. what do you think this means for creative freelancing? Are you guys connecting the same dots ?? submitted by /u/Legitimate_Aerie_606 [link] [comments]

two years ago this sub had 12k members asking "is claude better than chatgpt for writing" and now the company is worth a trillion dollars

I joined this sub when claude 3 opus dropped and it was a completely different world in here, small group of people who'd stumbled onto something that felt genuinely different from chatgpt and couldn't shut up about it. The posts were stuff like "did anyone else notice claude actually admits when it doesn't know something" and "i think anthropic might be onto something here" loll yesterday google committed $40 billion, amazon committed $25 billion the same week and revenue went from $1 billion to 30 billion in fifteen months which is apparently the fastest growth in american tech history. Secondary market says a trillion dollars and eight of the fortune 10 are customers, the tool we were geeking out about in a tiny subreddit is now arguably the most important ai product in the world and i'm still processing that I'm not trying to brag about being early because being early got me exactly nothing except a tool i love using and talk about too much at dinner parties. I'm writing this because i think this community deserves a moment and this sub was one of the first places where people figured out what claude could actually do in practice, people here were sharing creative pipelines, coding workflows and research systems openly before the enterprise market caught on. My own story is tiny compared to some of yours but it means everything to me, i do video content production and when i found this sub someone here posted about using claude to redesign their creative workflow and i tried the same thing and ended up in a conversation where claude basically told me my problem wasn't my tools it was my architecture,it helped me audit everything i was paying for separately across runway, topaz, heygen, kling, a headshot tool i used twiceand consolidate most of it into magichour, then connect the pipeline to remotion for automated editing. That single conversation saved me roughly $120 a month and cut my production time by 40%. I went from billing $3k a month doing everything manually to $14k a month as a one person studio and claude was involved in almost every step of that growth But honestly my story isn't the pointm hundreds of people in this sub have stories like this and collectively those stories are part of why anthropic is where it is today, the use cases now generating $30 billion in revenue started as experiments shared in communities exactly like this one. The part of the news i care about most as a daily user isn't the valuation it's the 10 gigawatts of new compute capacity. Every single person in this sub has hit rate limits midthought and wanted to throw something, if $73 billion in combined investment means i stop seeing "you've reached your limit" during a client deadline then the entire deal is justified and i will personally write dario a thank you letter haha I m trying not to get ahead of myself about what this means long term because historically when startups become megacorps the product changes and not always for the better but right now in this moment i just feel grateful i found this tool and this community when i did what's your claude story, curious when you joined and what changed for you because i think today's a good day to share those

Released a Claude Code skill that drives a full ML fine-tuning platform end-to-end from one conversation

We released a Claude Code skill today that turns a conversation with Claude into a deployed fine-tuned SLM. Point Claude at a folder of production traces, answer a few questions, approve a handful of checkpoints, and come back to a managed-endpoint deployment (or download the weights and self-host). The skill drives the process and distil labs platform does the heavy lifting under the hood (platform is free for open-source projects — details at the bottom). Concrete walkthrough from our release: 327 noisy production traces from a restaurant-reservation agent (the kind of data you can't train on directly) turn into a fine-tuned Qwen3-1.7B for multi-turn tool-calling. Install ``` curl -fsSL https://cli-assets.distillabs.ai/install.sh | sh distil signup /plugin marketplace add https://github.com/distil-labs/distil-cli-skill /plugin install distil-cli@distil-cli-skill ``` Then point Claude at your traces. What the skill actually does When I asked Claude "what are we about to do?" at the start, it laid out the 5-step workflow without me prompting for any of it: ``` Hey Claude, we will be training an SLM for multi-turn function calling based on the production logs in scenario-2-noisy-labels/. Can you explain the process on a high level? ⏺ 1. Prep. Inspect the traces, then write a converted traces.jsonl (OpenAI message format), a job_description.json, and a config.yaml. Upload & clean. The platform filters low-quality traces and uses a committee of teachers to relabel noisy outputs. Teacher eval. Verify the teacher can actually solve the task before spending hours on training. Train. Teacher generates synthetic data, student learns from it. Analyze & deploy. Compare the tuned student against the original production model, then deploy. ``` Every expensive step has a checkpoint where Claude stops for approval. Training burns credits that aren't easily refundable, so the confirmation checkpoint looked like this: ``` ⏺ Step 8: training confirmation checkpoint. Here's what will run: base.task: multi-turn-tool-calling-closed-book base.student_model_name: Qwen3-1.7B base.teacher_model_name: zai.glm-5 trace_processing.convert_to_single_turn: false trace_processing.teacher_model_name: zai.glm-5 Duration: ~6 hours Cost: burns credits, not easily refunded once started Pipeline: teacher eval → synthetic data gen → student finetune Reply "go" to kick off, or tell me what to change first. lets go! ``` The verdict-before-training gate refuses to even reach that checkpoint if teacher eval can't clear 0.70 on the generated test set. In that case, the skill walks you through iterating the task description instead. The skill also knows task-specific configs. For multi-turn tool-calling, I didn't have to look up that convert_to_single_turn: false is required. Claude flagged it as part of a config sanity check. Every checkpoint leaves a structured markdown analysis report (original-model-analysis.md, teacher-eval-analysis-iter-1.md, training-analysis-iter-1.md). Git-committable, reviewable three weeks later when someone asks why you picked this teacher. What came out A Qwen3-1.7B fine-tuned on ~10k synthetic examples grounded in the noisy traces. Model LLM-as-a-Judge staged_tool_call Function match Qwen3-1.7B (base, untuned) 0.513 0.535 45/78 GLM-5 (744B teacher) 0.808 0.695 69/78 Qwen3-1.7B (tuned) 0.846 0.769 76/78 Deployment Managed OpenAI-compatible endpoint (one-line swap in existing OpenAI client code), or download weights + Modelfile for llama.cpp or vLLM. Skill drives either path. Why it works as a skill Most skills I've seen wrap a few CLI commands but this one is end-to-end: reads your data, writes custom scripts, orchestrates an external platform, interprets the results, and leaves artifacts behind that persist past the conversation. The pattern that worked: Knows the workflow end-to-end and walks you through it Catches edge cases by re-reading the platform's own docs mid-conversation Stops for explicit approval on expensive operations Leaves structured artifacts that outlast the conversation Caveats Training is ~6 hours per run and burns credits (not refundable once started, which is why the confirmation gate exists). Happy to dig into how the checkpoints work, the config-sanity-check logic, or what building a purpose-built skill looked like. submitted by /u/party-horse [link] [comments]

I Built a desktop app for generating LLM fine-tuning datasets — started it a week ago while learning FT

Hey, I've been building side projects with Claude Code for a few months, but I'm completely new to fine-tuning — started experimenting maybe a week ago. From day one I wanted a GUI for the dataset side of the workflow, so this desktop app grew alongside my very first FT attempts. I know there are similar apps out there, but I wanted something simple that non-technical users could run with open-source models end-to-end. To sanity-check whether the datasets were actually useful I fine-tuned Qwen2.5-Coder-7B-Instruct on them and ran HumanEval / HumanEval+ (pass@1, 5 runs). Picked these benchmarks because they match the dataset's focus and run fast on my machine: I know it's not much but know now that app work :) - Base: 55.5% / 49.0% - FT V2 (1135 samples from the app): 60.0% / 54.0% Error bars don't overlap so it's at least not noise. Obviously HumanEval is only one slice — YMMV with other categories / criteria. https://reddit.com/link/1srz5aq/video/zubr426holwg1/player Stack: Next.js 16 + FastAPI + SQLite, packaged as standalone binary (Win/Linux). Code: https://github.com/AronDaron/dataset-generator Fine-tuned model: https://huggingface.co/AronDaron/Qwen2.5-Coder-7B-Instruct-DatasetGen-v2 Datasets: https://huggingface.co/datasets/AronDaron/dataset-gen-v1 / https://huggingface.co/datasets/AronDaron/dataset-gen-v2 Happy to hear feedback, especially if something doesn't work on your setup or if the approach misses something obvious — this is my first finetune llm tool release. submitted by /u/AronSan [link] [comments]