#humaneval

• Qwen 3.6 27B BF16 vs Q4_K_M vs Q8_0 GGUF evaluation (www.reddit.com) +9035 5d

  Evaluated Qwen 3.6 27B across BF16, Q4_K_M, and Q8_0 GGUF quant variants with llama-cpp-python using Neo AI Engineer. Benchmarks used: HumanEval: code generation HellaSwag: commonsense reasoning BFCL: function calling Total samples: HumanE…

  ↯ Qwen 3.6↯ Function Callinghumanevalfunction-callingllama+1

• Luce DFlash: Qwen3.6-27B at up to 2x throughput on a single RTX 3090 (www.reddit.com) +5015 6d

  Hey fellow Llamas, your time is precious, so I'll keep it short. We built a GGUF port of DFlash speculative decoding.

  ↯ Qwen 3.6humaneval

• "Second Thoughts" Been playing with adding a small transformer that reads output near the end of generation, and feeds it back near the top as a refinement loop. A quick test of 1.7B model showed drastic improvement in focused tasks (like coding) (bigattichouse.medium.com via reddit) +232 3h

  A 1.7B model can actually turn out some code, so I'm running the training for a 9B model, then will re-run HumanEval (a full one this time). I've shown most of my homework in the article, but will be posting to github after I clean things…

  humaneval

• I Built a desktop app for generating LLM fine-tuning datasets — started it a week ago while learning FT (www.reddit.com) +31 12d

  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…

  ↯ Qwen 2.5humanevalfine-tuningclaude-code

• Hito 2B: +35 on GSM8K, 75% on ARC-Challenge, 95% on HumanEval-style (www.reddit.com) +21 11d

  ↯ Qwen 3.5humaneval

• BigCodeBench: The Next Generation of HumanEval (huggingface.co) 97w

  humaneval

← all tags