Qwen3.8-27B ThinkingCap Abliterated (Merged)

SFT-merged LoRA (checkpoint-42, epoch 1) from the ThinkingCap efficient-reasoning dataset (hotdogs/thinkingcap-sft-qwen38-27b).

• Base: hotdogs/Qwen3.8-27B-abliterated (Qwen3_5, BF16, 55.6 GB)
• LoRA: r=32, alpha=64, target = all linear + GDN in_proj layers
• Training: 339 SFT rows, 1 epoch, max_seq 4096, 7x RTX 3090, BF16
• Eval loss (epoch 1): 0.2143 (best of 3 tracked epochs)
• MTP: 15 tensors preserved in blk.64 (restored post-merge)
• Dataset: hotdogs/thinkingcap-sft-qwen38-27b (369 SFT + 244 DPO pairs, on-policy from live Qwen3.8-27B-Ablit, oracle-verified 84.2%)
• GGUF versions: hotdogs/Qwen3.8-27B-thinkingcap-abliterated-mtp-GGUF

Files

19 files, 1199 tensor keys, 55.56 GB total.

Usage (transformers)

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
    "hotdogs/Qwen3.8-27B-thinkingcap-abliterated",
    torch_dtype="auto", device_map="auto", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("hotdogs/Qwen3.8-27B-thinkingcap-abliterated")

---

📊 Benchmark — real A/B data, not a guess

Every number below comes from live inference runs (same harness, same prompts, same flags, temperature=0, averaged over multiple runs). Three models were tested on the identical benchmark so the deltas are apples to apples:

Two suites were run:

• Easy — 52 single-shot items (multiple-choice, short math, puzzles), averaged over 5 runs (260 scored items per model).
• Hard — 10 multi-step algebra problems (same distribution as the ThinkingCap SFT data), averaged over 3 runs (30 items per model). This is where the models are pushed to think long enough to "run away."

Easy suite (5-run average, 260 items each)

Reading it: on short, easy prompts the thinking length is already small (~87–101 tokens) and all three land within each other's noise band. Accuracy is flat (≈94%). The cap is essentially free here — it costs nothing, which is exactly what you want from an efficiency layer.

Hard suite (3-run average, 30 items each)

Reading it: on hard multi-step problems the base model starts to run away (mean 636 chars, pulled up by one 4,868-char outlier). Abliteration is the biggest reducer (mean 636 → 235, −63%) and it holds accuracy. The cap sits between the two (358) — see the per-item breakdown below for why.

Per-item thinking length — Hard suite (chars, 3-run avg)

The pattern is the whole story:

• On the runaway item the cap does its job — it cuts the 4,868-char base spiral down to 1,843 (−62%), and abliteration cuts it further to 852. This is the real-world win: the cap stops the model from over-thinking.
• On short items (120–220 chars) all three are within a few percent — there is simply little to cut.
• On a couple of items (Sylvia, rectangle) the cap runs longer than abliteration. Abliteration removed the model's shortcuts, so it reasons more; the cap trims the runaway but leaves that extra reasoning in.

Verdict

ThinkingCap is runaway insurance, not a universal shrinker. It does not make every answer shorter — on short prompts it is a wash. What it reliably does is stop the model from spiraling on hard, multi-step problems (−62% on the runaway item) while keeping accuracy identical (70% on hard, ≈94% on easy, refusal still 0%). Abliteration is the bigger reducer; the cap is the guardrail on top. Stacked together (abliterated + cap = this model) you get the shortest typical thinking and the smallest worst-case thinking, at no accuracy cost.

How to reproduce

# bench.py is a self-contained stdlib-only harness (no pip install).
# Run the same command against each model on its own llama.cpp endpoint,
# then compare the JSON summaries. Easy suite:
python3 bench.py --api http://<ip>:8081 --suite all \
  --output base_normal --workers 4 --temp 0 --timeout 180
# ... x5 per model
# Hard (algebra) suite:
python3 bench.py --api http://<ip>:8081 --suite thinkhard \
  --output nh_cap --workers 4 --temp 0 --timeout 240
# ... x3 per model
# A/B two outputs:
python3 compare.py base_normal.json nh_cap.json

Flags kept identical across all runs: --workers 4 --temp 0, --timeout 180 (easy) / 240 (hard). Thinking length is reported as reasoning_content character count and completion_tokens from the API.

📊 Benchmarks

Tested with inspect_ai 0.3.260 on local llama.cpp, reasoning_effort=low, temperature=1, top_k=20, top_p=0.95. All runs are single-epoch (no multi-epoch averaging).

¹ Official vendor-reported scores from the Qwen3.8-27B model card, reasoning_effort=xhigh. ² Standard Google IFEval via inspect_evals/ifeval (541 prompts, prompt-level strict accuracy). ⁴ inspect_evals/swe_bench_verified_mini: 50 issues (25 django + 25 sphinx), message_limit=100 steps, max_tokens=16384/turn, tool_timeout=210s, no internet, Docker sandbox. ⁶ External/independent evaluations (not the official model card). SWE-bench Verified (500) is a different, harder split than our 50-task mini. Treat as directional, not exact. ⁷ External comparison via serenitiesai.com (F16 reference values). Treat as directional. ⁸ From SmartScope Gemini 3.1 Pro Benchmark Analysis (2026). GPT-5.5 and Kimi K3 SWE scores from MorphLLM SWE-bench Pro Leaderboard. ⁹ 31/49 of-scored (1 out: sphinx-doc__sphinx-9229, step-limited at 117 msgs). Run status error (APITimeoutError at scheduler level) but all 50 samples completed in summaries. ¹⁰ 34/49 of-scored (1 out: sphinx-doc__sphinx-9229, timeout at 162 msgs / 190 min). Run status cancelled — last sample hit the 3600s timeout.

Reading it: Q6_K scores within 2.1 pp of the full BF16 Qwen3.8-27B on GPQA (87.1 vs 89.2) — with a quantized, abliterated, LoRA-merged GGUF running on consumer GPUs. On IFEval, Q6_K (88.2%) and Q4_K_M (88.4%) are near-identical and sit within ~3 pp of the F16 reference (~91.5% for Qwen3.8-27B, 95.0% for Opus 4.6). On SWE-bench, Q6_K leads at 69.4% (34/49) vs Q4_K_M 63.3% (31/49) — both timed out on the same hard issue (sphinx-9229); external runs of the full 500-task SWE-bench Verified put Qwen3.8-27B at ~82% (footnote ⁶).

⚠️ Hardware limitation: Two-box setup — Box A: 1× RTX 3090 24GB + 1× RTX 3060 12GB (36 GB), Box B: 1× RTX 4060 Ti 16GB + 1× RTX 3060 12GB (28 GB). Total 64 GB VRAM across both boxes. F16 (55 GB) fits on Box A with tight KV-cache; Q6_K (21 GB) and Q4_K_M (16 GB) are the practical daily-drivers. F16 / Q8_K benchmark runs are pending — this table will be updated when they land.

⚠️ Token & step limits: SWE-bench is the most constrained benchmark: message_limit=100 (agent turns) × max_tokens=16384/turn ≈ 1.6M output-token ceiling. The model is typically step-limited, not token-limited (agent runs out of turns before exhausting the token budget). GPQA is single-shot with max_tokens=65536, so the token ceiling is rarely hit (avg ~4K tokens/question).

Smoke test (F16 GGUF)

• 27x43 = 1161 (think 163 chars)
• bat-and-ball = 0.05 (think 161 chars)
• snail wall = day 5 (think 359 chars, base was 4000+)