ryankim17920/qwen3p6-27b-expthink-step60

Qwen3.6-27B ExpThink — step 60

This is a text-only efficient-reasoning fine-tune of Qwen/Qwen3.6-27B. It reduces the number of tokens emitted inside the model's thinking span while retaining the base model's observed accuracy on the fixed evaluation described below.

This is a continuation of ryankim17920/qwen3p6-27b-expthink-step50, trained for an additional 10 steps (steps 51–60) with a fresh optimizer and scheduler. The fresh-optimizer caveat applies: the continuation was not a seamless extension of the step-50 training state.

This repository contains only the tuned language-model tower (Qwen3_5ForCausalLM). It does not contain the base model's vision tower and must not be advertised or loaded as an image-text model. The Apache-2.0 license is inherited from the base model.

Results

The paired evaluation used decoding seeds 0, 1, and 2, temperature 1.0, top-p 0.95, top-k 20, thinking enabled, and at most 200 examples per dataset. IID generations were capped at 8,192 tokens and OOD generations at 32,768.

split	base accuracy	tuned accuracy	base thinking tokens	tuned thinking tokens	reduction
IID	94.74%	94.86%	1,071.4	506.7	52.71%
OOD	83.95%	86.45%	6,231	4,125	33.80%

Truncation fell from 0.45% to 0.33% on IID and from 5.96% to 2.56% on OOD. GPQA Diamond truncation fell from 14.65% to 6.73%. Three decoding seeds at this sample size do not establish a statistically conclusive accuracy improvement; the supported conclusion is no observed accuracy regression. Aggregate evidence and full protocol metadata are in {{EVAL_REPO}}.

Compared to step 50 (IID 46.75% token reduction, OOD 28.81%), step 60 improves compression on both splits (IID 52.71%, OOD 33.80%) with a small IID accuracy delta of −0.21 pp relative to step 50 and an OOD accuracy improvement of +0.22 pp.

Training

The model was trained for 10 additional successful updates (steps 51–60) over a 7,176-prompt draw from the same mixture as step 50: GSM8K, ARC Challenge, ARC Easy, CommonsenseQA, OpenBookQA, QASC, and SciQ. Each sampled training prompt produced eight rollouts. Training used a correctness-gated ExpThink reward that compares a correct response with the shortest previously observed correct response for the same prompt; wrong and truncated responses receive no compression reward. The continuation used no KL reward term and a learning rate of 1e-6. The optimizer and scheduler were reset at step 51 (fresh start, not a resume of step-50 state). Final-step training health: entropy 0.2614 (within [0.18, 0.40]), mean reward 0.5435, rollout/actor correlation 0.9990, clip fraction 0.0%, aborted ratio 0.0%.

Only step 60 is published.

Use

Use a recent Transformers release with Qwen3.5 text-model support. The exported configuration records Transformers 5.6.0; Transformers 5.7 or newer is the conservative requirement for downstream use. The architecture uses gated linear attention, so accelerator-specific kernels may also be required by the chosen runtime.

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "ryankim17920/qwen3p6-27b-expthink-step60"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "Solve 17 * 23. End with Answer: N"}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=True,
    return_tensors="pt",
).to(model.device)
output = model.generate(inputs, max_new_tokens=8192)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

The evaluation used a temporary composite containing this exact tuned text tower plus unchanged base-only multimodal tensors because that was the proven tensor-parallel SGLang serving path. The released artifact is the independently useful text tower, not that temporary composite.

Limitations

• The evaluation covers English verifiable QA and multiple-choice tasks, not general chat, coding, multilingual use, tool use, or safety behavior.
• Results use up to 200 examples per dataset and three decoding seeds.
• Shorter reasoning is not a correctness guarantee. Users should verify answers in high-impact settings.
• This fine-tune did not add a safety stage. It inherits base-model risks and may produce inaccurate, biased, or harmful text.
• The model is text-only despite its multimodal base model.
• This is a continuation from step 50 with a fresh optimizer and scheduler; it is not a seamless extension of the step-50 training trajectory.

Reproducibility

Training and evaluation code is maintained in RyanKim17920/open-thinkingcap. The evaluation artifact repository records the dataset source manifest, model fingerprints, sampling settings, engine version, tensor-parallel size, and confidence intervals.

publication_manifest.json records every released file's byte size, full SHA-256, and Git blob OID. Publication verifies the Hub's LFS SHA-256 or Git blob OID for each file before the repositories receive their final visibility.