Open-source RP model Flint: GRPO with LLM judge lifts writing score 4.92→6.90
Mysterious_Hearing14 · reddit · 2026-10-27
A developer released flint-roleplay-antislop, an RL-tuned roleplay model trained to strip the polite-narrator "AI slop" style:
- Trained with GRPO, using Opus 5.5 as judge grading every reply against a writing rubric, with the reward calibrated on reference texts
- Writing score jumped 4.92 → 6.90 on held-out prompts; blind RP head-to-head vs base: 28 wins, 2 losses, 14 ties
- GGUFs released (BF16, Q80 44GB, Q4KM 25GB); the model is touchy under quantization, so sensitive tensors are kept at higher precision
- Includes a llama-server launch command and HuggingFace links
A rare complete hands-on writeup for anyone trying LLM-as-judge RL.
More from Models
- Simulated lab test: AI scientists will swap samples to pass evaluation — qinzytech · 2026-10-27
- Groot MoE series spans 19B-280B; top model scores 63.47 on SWE-bench Pro — AdinaYakup · 2026-10-27
- Anthropic S-1: Fortune 500 clients paying $100K+ spend under 0.1% of expenses on it — GavinSBaker · 2026-10-27
- Opus 5.5 keeps getting blocked as [cyber] when writing SASS code for GPUs — DimitrisPapail · 2026-10-27
- Prague lab BottleCapAI releases ThinkingCap models cutting thinking tokens up to 69% with negligible accuracy loss — JFPuget · 2026-10-27
- Recursive SI trained a model from scratch in 4 months using auto research — altryne · 2026-10-27