Uploaded October 2026 | Updated November 2026, 1 day ago
Can RL Make a 4B Model Compete with 27B? My Open RL Environment Experiments (GRPO, LoRA, DGX)
In this video I test whether open-source RL environments can make a small model perform closer to a much larger one by training Qwen 3.5 4B (with a LoRA) against Kaggle-style “small data” RL tasks on two DGX Sparks and comparing to Qwen 27B on a held-out exam. I explain what RL environments are (sandbox, tools, checker), why you need a strict checker to avoid reward hacking, and how GRPO works and only learns from tasks the model sometimes gets right. After weak gains from random tasks, I filter for “edge” tasks and get clear improvement, closing over half the gap, mostly by teaching the model to finish and answer correctly. Further RL hits a ceiling, so I try distillation from 27B, then RL on top, which gives the best results on unseen datasets. I share takeaways and put code, checker, and recipes on GitHub.
Thanks to FireCrawl for sponsoring this video:
Firecrawl, the context API for AI agents (free tier to get started)
firecrawl.link/promptengineer
💻 Code: harness, strict checker, training scripts and results
github.com/PromtEngineer/smoldataenvs-grpo-spark
📚 Resources
SmolDataEnvs (the RL environments): huggingface.co/datasets/FineEnvs/SmolDataEnvs
FineEnvs: github.com/adithya-s-k/FineEnvs
Qwen3.5-4B: huggingface.co/Qwen/Qwen3.5-4B
Qwen3.5-27B: huggingface.co/Qwen/Qwen3.5-27B
TRL GRPO Trainer: huggingface.co/docs/trl/grpo_trainer
NVIDIA DGX Spark: nvidia.com/en-us/products/workstations/dgx-spark
Xiaomi MiMo-V2.6 (open RL environments): mimo.mi.com/docs/en-US/news/latest/v2-6
TechCrunch on RL environments: techcrunch.com/2025/09/21/silicon-valley-bets-big-on-environments-to-train-ai-agents
Website: engineerprompt.ai
RAG Beyond Basics Course:
prompt-s-site.thinkific.com/courses/rag
Let's Connect:
🦾 Discord: discord.com/invite/t4eYQRUcXB
☕ Buy me a Coffee: ko-fi.com/promptengineering
|🔴 Patreon: patreon.com/PromptEngineering
💼Consulting: calendly.com/engineerprompt/consulting-call
📧 Business Contact: engineerprompt@gmail.com
Become Member: tinyurl.com/y5h28s6h
💻 Pre-configured localGPT VM: bit.ly/localGPT (use Code: PromptEngineering for 50% off).
Signup for Newsletter, localgpt:
https://tally.so/r/3y9bb0
00:00 Can RL Shrink the Gap
01:24 What RL Environments Are
02:45 Reward Hacking and Checkers
03:37 Where RL Fits in Training
04:32 GRPO and Learning Signal
05:17 Hardware Setup and Baselines
06:18 Sponsor Firecrawl
07:51 Training on Edge Tasks
09:39 Before and After Runs
11:37 Why Progress Stalls
12:38 Distillation from 27B
13:51 Best Combo Model Example
14:52 Takeaways and Wrap Up
Can RL Make a 4B Model Compete with 27B? My Open RL Environment Experiments (GRPO, LoRA, DGX)
In this video I test whether open-source RL environments can make a small model perform closer to a much larger one by training Qwen 3.5 4B (with a LoRA) against Kaggle-style “small data” RL tasks on two DGX Sparks and comparing to Qwen 27B on a held-out exam. I explain what RL environments are (sandbox, tools, checker), why you need a strict checker to avoid reward hacking, and how GRPO works and only learns from tasks the model sometimes gets right. After weak gains from random tasks, I filter for “edge” tasks and get clear improvement, closing over half the gap, mostly by teaching the model to finish and answer correctly. Further RL hits a ceiling, so I try distillation from 27B, then RL on top, which gives the best results on unseen datasets. I share takeaways and put code, checker, and recipes on GitHub.
Thanks to FireCrawl for sponsoring this video:
Firecrawl, the context API for AI agents (free tier to get started)
firecrawl.link/promptengineer
💻 Code: harness, strict checker, training scripts and results
github.com/PromtEngineer/smoldataenvs-grpo-spark
📚 Resources
SmolDataEnvs (the RL environments): huggingface.co/datasets/FineEnvs/SmolDataEnvs
FineEnvs: github.com/adithya-s-k/FineEnvs
Qwen3.5-4B: huggingface.co/Qwen/Qwen3.5-4B
Qwen3.5-27B: huggingface.co/Qwen/Qwen3.5-27B
TRL GRPO Trainer: huggingface.co/docs/trl/grpo_trainer
NVIDIA DGX Spark: nvidia.com/en-us/products/workstations/dgx-spark
Xiaomi MiMo-V2.6 (open RL environments): mimo.mi.com/docs/en-US/news/latest/v2-6
TechCrunch on RL environments: techcrunch.com/2025/09/21/silicon-valley-bets-big-on-environments-to-train-ai-agents
Website: engineerprompt.ai
RAG Beyond Basics Course:
prompt-s-site.thinkific.com/courses/rag
Let's Connect:
🦾 Discord: discord.com/invite/t4eYQRUcXB
☕ Buy me a Coffee: ko-fi.com/promptengineering
|🔴 Patreon: patreon.com/PromptEngineering
💼Consulting: calendly.com/engineerprompt/consulting-call
📧 Business Contact: engineerprompt@gmail.com
Become Member: tinyurl.com/y5h28s6h
💻 Pre-configured localGPT VM: bit.ly/localGPT (use Code: PromptEngineering for 50% off).
Signup for Newsletter, localgpt:
https://tally.so/r/3y9bb0
00:00 Can RL Shrink the Gap
01:24 What RL Environments Are
02:45 Reward Hacking and Checkers
03:37 Where RL Fits in Training
04:32 GRPO and Learning Signal
05:17 Hardware Setup and Baselines
06:18 Sponsor Firecrawl
07:51 Training on Edge Tasks
09:39 Before and After Runs
11:37 Why Progress Stalls
12:38 Distillation from 27B
13:51 Best Combo Model Example
14:52 Takeaways and Wrap Up







![Training Your Own Embedding Model Is Not As Hard As You Think
Fine-tune Embedding Gemma 2 (EmbeddingGemma 2) on your own data with Unsloth and LoRA in a free Colab notebook. I show how fine-tuning an embedding model differs from fine-tuning an LLM, then train Googles multimodal embedding model on two tasks: environmental sounds (ESC-50 top-1 24.5% to 65.8%) and search over my own YouTube transcripts (66.3% to 75.0% top-1 on videos it never saw), and check what it costs the rest of the model.
Youll learn how to build question-passage training pairs, why in-batch negatives (MultipleNegativesRankingLoss) need a no-duplicates sampler, why training and search prompts must match, how LoRA works on a shared multimodal backbone, how to evaluate on unseen data, and a reload bug that silently breaks audio adapters.
If you have fine-tuned embedding models for your own applications, share what worked for you in the comments.
Links:
Colab notebook: https://colab.research.google.com/drive/1xq-84DLZsqteEzL52Qb 47Y_yGckeyB?usp=sharing
The first EmbeddingGemma 2 video: [EMBEDDINGGEMMA 2 VIDEO LINK]
Unsloth EmbeddingGemma 2 guide: https://unsloth.ai/docs/models/embeddinggemma-2
Unsloth notebooks: https://github.com/unslothai/notebooks
Model card: https://huggingface.co/google/embeddinggemma-2
ESC-50 dataset: https://github.com/karolpiczak/ESC-50
Sentence Transformers losses: https://sbert.net/docs/package_reference/sentence_transformer/losses.html
00:00 - Fine-Tuning EmbeddingGemma 2 on Your Own Data
01:03 - Embedding Fine-Tuning vs LLM Fine-Tuning
02:30 - In-Batch Negatives (MultipleNegativesRankingLoss)
03:37 - The Duplicate Trap: No-Duplicates Sampler
04:45 - LoRA on a Shared Multimodal Backbone
06:30 - Notebook: ESC-50 Sounds with Unsloth on an A100
09:32 - Side Effects on Photo, Voice and Text Search
10:06 - Gotcha: Reloading an Audio Adapter
10:50 - Fine-Tuning on YouTube Transcripts
12:04 - Three Things to Get Right Training Your Own Embedding Model Is Not As Hard As You Think](https://i.ytimg.com/vi/S7tFyREI19I/mqdefault.jpg)


