Show HN: Tiny 1B param model that beats GPT-3.5 on JSON extraction
Quick Answer
A newly trained 1B parameter Llama-3 derivative outperforms GPT-3.5 in JSON extraction tasks, achieving superior results on a 10-task benchmark.
Quick Take
A newly trained 1B parameter Llama-3 derivative outperforms GPT-3.5 in JSON extraction tasks, achieving superior results on a 10-task benchmark. It runs at 80 tokens per second on a single 4090 GPU, with both weights and evaluation suite available as open-source.
Key Points
- Model trained on 200K synthetic JSON-extraction examples.
- Outperforms GPT-3.5 on a 10-task held-out benchmark.
- Achieves 80 tokens per second on a single 4090 GPU.
- Weights and evaluation suite are open-sourced for public use.
- Significant implications for developers working with JSON data.
Article Excerpt
From source RSS / original summaryI trained a 1B parameter Llama-3 derivative on 200K synthetic JSON-extraction examples. It beats GPT-3. 5 on a 10-task held-out benchmark while running at 80 tok/s on a single 4090. Weights and eval suite open-sourced.
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from Hacker News
See more →Show HN: RLM-based local debugger for AI agent traces
HALO (Hierarchal Agent Loop Optimizer) is an open-source tool designed for debugging AI agents by analyzing OTEL compliant execution traces. It utilizes a Recursive Language Model (RLM) to efficiently identify patterns and systemic issues, enabling developers to optimize their agents iteratively without complex setups.
