AI/ML●●Solid
Flint – A 30B model fine-tuned for less repetition
Fine-tuned Qwen 30B that prioritizes output diversity over convergent accuracy.
Niche GemSolve My Problem
thmsmxwll
622mo ago
Back to browse
Fine-tuned 3B outperforms Claude Haiku on constrained generation
by serendip-ml·Mar 13, 2026·1 point·0 comments
AI Analysis
●●SolidBig BrainNiche Gem
Fine-tuned 3B Qwen matches Haiku on jokes, validating small models for constrained agent tasks.
Strengths
- •Concrete benchmark data comparing 0.5B-72B Qwen models against Claude Haiku 4.5.
- •Details SFT and DPO pipeline using 3000 samples, reproducible for other tasks.
- •Challenges RAG-heavy agent architecture by showing trained adapters outperform prompts.
Weaknesses
- •No model weights or code repository linked for immediate implementation or testing.
- •Joke telling is a narrow benchmark; results may not generalize to complex reasoning.
Category
Target Audience
ML Engineers, AI Agent Developers
Similar To
HuggingFace · Axolotl · Unsloth
Similar Projects
AI/ML●●Solid
I fine-tuned Qwen 3.5 (0.8B–4B) on a Mac for text-to-SQL – 2B beats 12B
Unified memory trick lets a 2B model beat 12B; trains on MacBook with zero cloud costs.
Ship ItNiche GemBig Brain
sciences44
713mo ago
AI/ML●Mid
100% LLM accuracy–no fine-tuning, JSON only
Ancient Rome Q&A benchmark shows 81pp accuracy lift, but lacks adversarial defense evidence.
Big Brain
MysticBirdie
223mo ago
AI/ML●Mid
I fine-tuned Gemma 4 to talk like a pirate
Cool demo, but there's no actual tool to use—just a video and writeup.
Ship It
logicallee
302mo ago
AI/ML●●Solid
Shard-based scheduling for 100x more fine-tuning experiments on 4 GPUs
Shard-based scheduling cuts GPU wait time, though Ray Tune offers similar early stopping.
Big BrainSolve My Problem
kamranrapidfire
102mo ago
AI/ML●Mid
OpenAI CLIP fine tuned on Galaxy morphology
Galaxy classification model, but model card has mostly empty fields.
Niche Gem
mjupp1
102mo ago