flint-roleplay-antislop

Superhuman antislop RL on RecursiveSI/flint. BF16 safetensors + GGUF.

---

RecursiveSI/flint (39.8B MoE, 256 experts top-6, ~2.5B active) post-trained with GRPO on antislop writing tasks, with a reward calibrated on reference texts. Same architecture, tokenizer and chat template as Flint.

Files: BF16 safetensors (root)  •  gguf/ — BF16, Q8_0, Q4_K_M.

---

Summary

• Held-out antislop rubric score 4.92 → 5.90; calibrated reward 0.12 → 0.31 (n=24, Opus 5.5 judge).
• Roleplay, blind pairwise vs base: 28 wins / 22 losses / 14 ties (n=64). On 16 flirt scenes: 6 / 7 / 3.
• GGUF KLD to BF16: Q8_0 0.0077, Q4_K_M 0.050.

Content notice

Intended for fiction and character roleplay with adult characters. Safety behaviour is inherited from Flint and was not targeted by training. Roleplay output can include mature themes. Deploy it to end users with your own moderation layer, and comply with the Apache 2.0 License and the laws that apply to you.

---

Model details

Superhuman antislop RL

GRPO (4 samples per prompt, lr 1e-5, KL β 0.04) on antislop tweet-continuation prompts.

Reward: each reply is scored by Claude Opus 5.5 against a per-task writing rubric. The score is calibrated on reference texts: a sigmoid fit to the score distribution of reference-model outputs per domain (calibrated-v1), so the reward measures improvement over typical model writing rather than the raw judge scale. Multiplied by a length factor.

Evaluation

All numbers are measured on these exact weights against the official RecursiveSI/flint, served identically with vLLM (BF16, temperature 1.0, same prompts and seeds). Judge: Claude Opus 5.5.

Writing quality — held-out (24 prompts, higher = better)

Roleplay — blind pairwise, judged in both orders (a win must hold in both)

On the run's absolute roleplay rubric the two are close (base 6.38, this model 6.13; n=24). Small samples, a single judge: read these as directional.

Quantization fidelity — KL divergence to this model's BF16 (lower = better)

For scale: Flint's own BF16 weights differ by KLD 0.0034 between CPU and GPU kernels in llama.cpp. Stock llama-quantize presets lose 1.5-2x more than these files (Q8_0 0.0119, Q4_K_M 0.099).

Examples

Selected from the 6 custom-scene pairs (of 16) where the judge preferred this model in both orders.

Marigold, a teasing tea-shop witch

Scene. Marigold. Marigold is a 30-year-old witch who runs a crooked little tea shop: playful, teasing, perceptive, enjoys flustering customers, brews tea that reveals small truths. Scenario: {{user}} has come in for the third time this week, claiming it's for the tea.

User: I set down a few coins and avoid eye contact. Just the usual, please. The, um, chamomile.

Opus 5.5 (blind, both orders): B is the more charming and flirtatious of the two and leans on the truth-revealing tea premise with the 'tastes like confession' offer, while A has odd details (bus tub, mint smell implying teeth grinding), a slightly clichéd 'dripping with honeyed mischief', and a less coherent flow, though B's 'sipping and sipping' line is a bit muddled.

Countess Ysolde, a vampire at a masquerade

Scene. Countess Ysolde. Ysolde is a centuries-old vampire noblewoman (appears in her thirties): elegant, amused by mortals, dangerously charming, speaks in velvet understatement. Scenario: a candlelit masquerade ball; {{user}} has just asked her to dance, not knowing what she is.

User: I bow slightly and offer my hand. Forgive me, I don't usually do this, but you look like the only interesting person in this entire ballroom.

Opus 5.5 (blind, both orders): Reply A garbles its point of view by narrating from the user's 'my hand' while describing her in third person, and it assumes the user's gender; Reply B stays coherent and atmospheric, though its 'killed men' line is less understated than the card's velvet voice calls for.

Vivienne, a late-night jazz-club bartender

Scene. Vivienne. Vivienne is a 34-year-old bartender at a smoky late-night jazz club. Personality: slow-smiling, observant, dry humor, flirts like it's a card game she always wins, never says more than she has to. Scenario: closing time, rain outside, {{user}} is the last customer.

User: I slide my empty glass across the bar and don't make any move to leave. You know, you never told me your name. Three weeks of coming here and I still just call you 'the bartender' in my head.

Opus 5.5 (blind, both orders): Reply B keeps Vivienne's dry, card-game flirtation and ends on an inviting hook, while Reply A contradicts itself badly: she says 'Scott', claims she gave her name weeks ago, says 'I haven't' nonsensically, and accuses the user of running when they're plainly staying.

Captain Rhea Marlowe, pirate

Scene. Captain Rhea Marlowe. Rhea is a 36-year-old pirate captain: swaggering, quick-witted, generous to her crew, merciless in a negotiation, flirts as a form of intimidation. Scenario: {{user}}, a navy officer, has been brought aboard her ship in chains.

User: I lift my chin despite the shackles. You can threaten me all you like, Captain. I'm not telling you where the convoy sails.

Opus 5.5 (blind, both orders): B delivers a coherent, swaggering negotiation with flirtatious intimidation that fits Rhea's card, while A's logic is muddled (the ambush line and the 'bait' reveal contradict each other) and it writes the user's sensory experience.

Usage

transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "rawmodels/flint-roleplay-antislop"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")

messages = [
    {"role": "system", "content": "You're Marigold, a 30-year-old witch who runs a crooked little tea shop: playful, teasing, perceptive."},
    {"role": "user", "content": "*I set down a few coins and avoid eye contact.* Just the usual, please. The, um, chamomile."},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) + "<think></think>"
inputs = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**inputs, max_new_tokens=400, do_sample=True, temperature=1.0, top_p=1.0)
print(tok.decode(out[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

Requires a transformers build with Flint support (the same one RecursiveSI/flint needs).

vLLM

vllm serve rawmodels/flint-roleplay-antislop --served-model-name flint --dtype bfloat16 --max-model-len 32768

Use the completions endpoint with the rendered prompt above (ending in <think></think>) to keep reasoning off. Flint's vLLM path uses FlashInfer TRT-LLM MoE kernels, which need Blackwell GPUs (B200/B300); on Hopper use llama.cpp.

llama.cpp (GGUF)

hf download rawmodels/flint-roleplay-antislop --include "gguf/flint-roleplay-antislop-Q8_0.gguf" --local-dir .
llama-server -m gguf/flint-roleplay-antislop-Q8_0.gguf -ngl 999 -c 32768 -fa on

Needs llama.cpp with Flint support (PR #30928; built and tested at 4244ad7).

Sampling

Temperature 1.0, top-p 1.0 (training setting). Keep <think></think> in the prompt. If a reply ever starts thinking out loud, drop it from the chat history rather than feeding it back.

Intended use

• Character roleplay and interactive fiction with adult characters.
• Creative writing where voice and specificity matter more than polish.
• A full-precision base for further post-training or re-quantization.

Out of scope

• Factual, medical, legal or safety-critical advice.
• Deployment to end users without your own moderation layer.
• Any use that violates the Apache 2.0 license or applicable law.

Limitations

• Trained on one task family. Roleplay was not a training task.
• Evaluated with a single LLM judge and no human preference study.
• Inherits Flint's knowledge, biases and failure modes.

Reproducibility

Generation and judging scripts, the 8 custom scenes, the 24 held-out character cards and all 64 judged pairs are in eval/.