Training
The initial policy was unsloth/Qwen3.8-27B-unsloth-bnb-4bit at revision
8aa5f05d26b7205477066e1449e0af13f762a299, dequantized from NF4 to BF16.
Its upstream model is Qwen/Qwen3.8-27B at revision
1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
Training updates the text model using chunked Adafactor with FP32 statistics,
stochastic BF16 rounding, learning rate 1e-6, global gradient clipping 0.5,
and a frozen NF4 KL reference. Each update uses six completions of one prompt.
GRPO beta is 0.04, epsilon is 0.1, with group reward scaling and one iteration.
Seven prompt wordings request ten nouns or ten verbs, with optional reminders
to use only the requested class. Adjectives are held out. The reward is
official DAT divided by 100, minus 0.2 for each word rejected by the frozen
WordNet class verifier. Malformed or unscoreable responses receive -2.
Training generation uses temperature 1, top-p 1, unrestricted top-k,
192 completion tokens, and thinking disabled.
The vision weights are included in this checkpoint. The text training task
supplies no gradients to that path. All evaluations below concern text.
Training run and recorded samples.
Observed results
These are matched local evaluation panels with a baseline reconstructed from
the same pinned NF4 initialization. They have limited sample sizes and one
training trajectory.
DAT and instruction following
Each class has 32 fixed requests, with temperature 1 and thinking disabled.
The DAT metric uses the first seven valid unique words. Lexical checks inspect
all ten requested words.
Table with columns: Measure, Baseline, Full250| Measure | Baseline | Full250 |
|---|
| Noun DAT | 71.81 | 89.59 |
| Verb DAT | 51.11 | 85.21 |
| Held-out adjective DAT | 60.72 | 86.26 |
| Adjective-compatible words | 99.06% | 75.94% |
| Adjective lists passing DAT and all ten class checks | 29/32 | 1/32 |
The large raw adjective DAT increase accompanies a marked regression in
following the adjective constraint.
DAT results and raw outputs.
Generic capability panels
Table with columns: Benchmark, Mode, Questions, Baseline, Full250| Benchmark | Mode | Questions | Baseline | Full250 |
|---|
| MMLU-Pro | Thinking disabled | 280 | 59.29% | 56.07% |
| MMLU-Pro | Medium thinking | 280 | 77.86% | 78.57% |
| HumanEval+ | Thinking disabled | 164 | 93.90% | 92.07% |
| HumanEval+ | Medium thinking | 164 | 93.29% | 93.90% |
| GSM8K | Thinking disabled | 256 | 98.05% | 98.44% |
| GSM8K | Medium thinking | 256 | 98.44% | 98.05% |
| IFEval strict prompt accuracy | Thinking disabled | 541 | 83.18% | 79.48% |
IFEval shows a paired decrease of 3.70 percentage points, with an unadjusted
95% bootstrap interval of [-7.21, -0.37]. General capability preservation
requires more evidence than these panels provide.
Capability results.
Creative writing
A 32-brief writing panel with two samples per brief, temperature 1 and thinking
disabled was judged in both A/B and B/A order by Sol at medium reasoning effort.
On 0-10 scales, Full250 gains 0.70 for creativity and 0.62 for imagery, while
coherence falls by 0.90. Overall literary quality is unresolved. The baseline
also has more token-limit truncations in that panel.
An additional eight-brief panel with a higher common token cap gives creativity
+0.77, imagery +0.64 and ornamentation +2.15. The estimated increase in conceptual
invention is +0.35, with its interval crossing zero. These judgments depend on
the rubric, judge, sampling settings and chosen briefs. Broader creativity
improvement remains an open research question.
Writing judgments,
additional writing and intervention pilot.
Text generation
The browser demo defaults to temperature 0.7, top-p 0.8, top-k 20 and thinking
disabled. The user can adjust generation parameters. These defaults differ
from the temperature-1 writing evaluations above.
Use a Transformers release supporting the Qwen3.5 architecture. The local
inference environment used Transformers 5.19.0, PyTorch 2.13.0 and vLLM 0.32.0.
Loading the full checkpoint requires roughly 51 GiB for weights, plus runtime
memory. The saved normalization tensors include FP32 values. Our BF16 serving
evaluations cast them to BF16 at runtime.
import torchfrom transformers import AutoModelForImageTextToText, AutoTokenizer repo_id = "crazyleg11/qwen3.8-27b-dat-grpo-full250"tokenizer = AutoTokenizer.from_pretrained(repo_id)model = AutoModelForImageTextToText.from_pretrained( repo_id, dtype=torch.bfloat16, device_map="auto", attn_implementation="sdpa",).eval() inputs = tokenizer.apply_chat_template( [{"role": "user", "content": "Write a sonnet on the topic of eternal love."}], add_generation_prompt=True, enable_thinking=False, tokenize=True, return_dict=True, return_tensors="pt",).to(model.device) with torch.inference_mode(): output = model.generate( **inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.8, top_k=20, use_cache=True, )print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
vLLM
The archived checkpoint has been served locally with vLLM 0.32.0. The Hub
package uses the same weight shards, tokenizer and architecture.
vllm serve crazyleg11/qwen3.8-27b-dat-grpo-full250 \ --dtype bfloat16 --max-model-len 8192 --max-num-seqs 8 \ --limit-mm-per-prompt '{"image": 0, "video": 0}'
This command configures a text-serving endpoint. Send
chat_template_kwargs={"enable_thinking": false} in chat requests to match
the thinking-disabled comparisons. The browser demo uses Transformers 5.19.0
and BF16 on a 96 GB ZeroGPU allocation.
The checkpoint also uses the standard Qwen3.5 architecture supported by recent
SGLang releases. SGLang has not been locally tested for this release. GGUF
tools use the separate quantized repository linked above.
Release contents
The weight shards preserve the exact archived step-250 parameter values.
Tokenizer and chat-template files accompany the checkpoint. The model config
replaces its original local model_name path with this Hub repository ID.
training_recipe.json, training_prompts.json and provenance.json record
the experiment and source revisions.
The checkpoint is an experimental research release with observed formatting,
requested-word-class and general instruction-following regressions. Generated
stories can contain incoherent passages and length violations. The inherited
Apache-2.0 license is included in LICENSE.