Qwen3.8-27B-antislop

Try it (Space) · GGUF · rawmodels.ai · Original paper: unslop / RL-XAR (Meta FAIR) · RL environment · Reproduce

Qwen3.8-27B post-trained to write less like an LLM and more like a person. #4 of 29 on EQ-Bench 4 (Elo 1320, within the CI of #2 and #3), above GPT-5.5 and Claude Opus 4.7/4.8. GRPO against calibrated expert-aligned rubrics, our take on Meta FAIR's unslop / RL-XAR.

---

Background: AI slop

Current LLMs write fluent imitations of good prose that match the surface: tweets that read like a brand account, essays that resolve into generic motivational endings, "mini-paper" abstracts that over-cover and dump numbers. Standard pairwise and single-score LLM judges fail to catch this — they often prefer model writing to real expert writing. Following FAIR's unslop work, the fix is rubrics that separate expert from model writing, used as an RL reward.

Method: calibrated expert-rubric reward + GRPO

Our take on RL-XAR, on a text-only prose slice:

• Rubrics. For each expert excerpt (tweets from the Broad Twitter Corpus, scientific abstracts from arXiv, blog essays from PG / Sivers / sama), a frozen per-item rubric is generated and kept fixed for the whole run — no rubric drift during RL.
• Length-blind judge. A calibrated Claude Opus judge scores a candidate against its rubric on a 0–10 quality scale q, with a length-blind system prompt so it rewards craft.
• Calibrated reward. reward = L(words, target) · σ(a_d · (q − μ_d)), where L is a length factor against the item's target length and (μ_d, a_d) are a per-domain calibration fit on the Qwen3.8-27B reference distribution. This keeps the signal about quality
• RL. GRPO on Qwen/Qwen3.8-27B via verl, batch 128 (16 prompts × 8 samples), 100 steps, lr 2e-5, kl_loss_coef 0.002, thinking disabled. Generation is served by vLLM colocated with FSDP.

Over 100 steps the mean reward rose ~0.35 → 0.55 while entropy settled from 1.5 to ~0.43 (convergence on a consistent voice).

Results

Held-out prose, graded by the same calibrated length-blind Opus judge on the same frozen rubrics. We compare the base model, this model (antislop), and Altworld/Hemmingway-1 (a creative-writing finetune of the same base).

Pairwise win-rate (per-item, higher q):

• antislop vs base: 50–0 (2 ties)
• antislop vs Hemmingway-1: 43–6 (3 ties) → 82.7%

EQ-Bench 4

Official leaderboard config (same personas, judge panel and anchors), folded into the anchor Elo table.

Measuring it on EQ-Bench 4

Qwen's chat template opens <think> inside the prompt, so the reply carries only a closing </think>. EQ-Bench strips <think>…</think> pairs and misses the orphan, and the persona and judges then read the reasoning. Serve the model with thinking off and a reasoning parser:

vllm serve AlexWortega/Qwen3.8-27B-antislop --served-model-name antislop \
  --reasoning-parser qwen3 --default-chat-template-kwargs '{"enable_thinking": false}' \
  --max-model-len 32768

Check that content has no </think> before you start:

curl -s localhost:8000/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"antislop","messages":[{"role":"user","content":"hi"}]}' | grep -c '</think>'   # expect 0

Then run EQ-bench/eqbench4 with the leaderboard panel, pointing antislop at http://localhost:8000/v1/chat/completions in config.yaml, and fold the result into the anchor table:

uv run python eqbench4.py --evaluated-model antislop \
  --persona-models google/gemini-3.1-pro-preview \
  --pairwise-judges google/gemini-3.1-pro-preview,openai/gpt-5.5,claude-opus-4-6 \
  --pointwise-judges claude-opus-4-8 \
  --run-id antislop --results-dir results --out results/antislop.json
uv run eqbench --redo-elo <the same arguments>   # rebuilds Elo against the leaderboard anchors

Examples — baseline vs antislop

Freshly generated from this model vs the base, same prompt, same decoding (enable_thinking=False, temp 0.8). The baseline reaches for abstract, generic phrasing and tidy clichés; the antislop model stays concrete and in-voice.

Continue an essay — "The problem with productivity advice is that it treats attention as a resource to be spent, when really"

base: …it's more like a weather pattern. You don't "manage" the rain… Stop trying to hoard every minute like it's gold. The quiet gaps aren't waste; they're where the actual thinking happens. antislop: …it's more like a mood. You just end up staring at it for forty minutes because the cursor is blinking and you're too tired to close the tab.

Continue a product memo — "We looked at why users churn in week one, and the honest answer is"

base: …they can't find value fast enough. The onboarding flow is too long… If we don't make the first ten seconds count, they'll leave before they even understand what the product does. antislop: …they don't know what to do next. The dashboard is empty, the onboarding tooltip disappears after three seconds, and there's no clear button saying "invite your team". They close the tab, assume it's broken, and never come back.

Finish a casual tweet — prefix "first coffee of the day and"

base: first coffee of the day and i'm finally functional. save me, caffeine. antislop: first coffee of the day and i already know i'm not functioning before 10am

What's in this repo

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("AlexWortega/Qwen3.8-27B-antislop")
model = AutoModelForCausalLM.from_pretrained("AlexWortega/Qwen3.8-27B-antislop",
                                             dtype="bfloat16", device_map="auto", trust_remote_code=True)
msgs = [{"role": "user", "content": "Finish this tweet naturally: Monday mornings are"}]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False)
ids = tok(text, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(**ids, max_new_tokens=64, do_sample=True, temperature=0.8)[0][ids.input_ids.shape[1]:]))

Training details & versions

Training and eval scripts are in training_code/; REPRODUCE.md runs them end to end.

Acknowledgements & citation

This is an independent reimplementation of Meta FAIR's RL-XAR / unslop idea on a different base model and a different judge/calibration stack. Please cite the original work:

@misc{fair_unslop,
  title  = {Unslop: RL from eXpert Aligned Rubrics (RL-XAR)},
  author = {Meta FAIR},
  url    = {https://facebookresearch.github.io/RAM/blogs/unslop/},
  note   = {Research blog}
}