qwen3.5-elt-l3

Qwen3.5 ELT L3

This repository publishes the L3 Qwen3.5 Elastic Looped Transformer research handoff as Hugging Face safetensors plus GGUF artifacts.

The strongest supported claim is narrow: this artifact is strong on the local STEM bridge task and loop depth helps in the loop-aware HF runtime. It is not yet a broad LLM benchmark or GSM8K/math-generalization claim.

Files

• model-00001-of-00007.safetensors ... model-00007-of-00007.safetensors - merged BF16 Hugging Face shards.
• model.safetensors.index.json - HF sharded-weight index.
• elt-lm-qwen35-side-stem-aha-ilsd-l3-BF16.gguf - BF16 GGUF handoff.
• elt-lm-qwen35-side-stem-aha-ilsd-l3-Q8_0.gguf - Q8_0 GGUF handoff.
• elt-lm-qwen35-side-stem-aha-ilsd-l3-TQ4_1S.gguf - TurboQuant-style offline weight-compression GGUF.
• elt_export_manifest.json and publish_manifest.json - export and publication metadata.

Runtime boundary

The GGUF files carry ELT loop metadata and are useful for structural and single-pass runtime handoff, but stock GGUF execution is not the loop-aware L>=2 runtime. The config records elt.gguf.runtime_status=requires_looped_qwen35_runtime; use the loop-aware Qwen3.5 HF runtime for L-depth quality experiments.

The TQ4_1S file is an offline weight-compression artifact. It is not a claim about TurboQuant KV-cache serving performance.

Evaluation snapshot

Two-sided Fisher exact p-values over correct/incorrect counts:

Loop-aware HF runtime quality uses paired case IDs on 32 local STEM bridge questions and MCQ log-probability scoring:

Paired exact McNemar/binomial result for L1 vs L3: 7 improved, 0 regressed, p=0.01562.

2026-05-20 KV/Triality evidence

The current K-protected TheTom KV smoke keeps K at q8_0 or bf16 and sweeps V through turbo3, turbo4, and a turbo8 parser probe. turbo3/turbo4 have successful local rows; turbo8 is not supported by the installed llama.cpp runtime (Unsupported cache type: turbo8) and is not a measured throughput or quality result.

Triality SO(8) vector-view audit over 3/4/8-bit rows passed the 0.01 gate:

Loop-aware paired CV over 32 local STEM bridge cases gives L1=0.4375, L2=0.5625, L3=0.6562 accuracy. Paired permutation p-values are 0.122488 (L1-L2), 0.016098 (L1-L3), and 0.254775 (L2-L3); Friedman within-block permutation p=0.002500.

A logged lm-eval-harness gate now scores 128 external heldout rows against the Q8_0 GGUF through llama-server --ngl 999: 64 native MMLU-STEM MCQ rows plus 64 GSM8K test rows converted into numeric multiple-choice questions. The run uses the harness evaluator with a local llama.cpp /completion log-prob adapter because this installed server exposes the newer logprob schema. It is not a broad leaderboard result or standard GSM8K exact-match claim.

The four-policy Friedman within-fold permutation p-value is 0.781022. Pairwise paired within-fold permutation p-values are 1.000000 except the two bf16,V=turbo3 vs V=turbo4 comparisons, which are 0.750973. MMLU-STEM is 0.7031 accuracy for all four measured policies; GSM8K numeric-MCQ ranges from 0.3906 to 0.4219. turbo8 remains an unsupported runtime/parser value in this llama.cpp build.

The publication audit is included as eval/goal_completion_audit.md and eval/goal_completion_audit.json. It records the same claim boundaries: turbo8 is unsupported in this runtime, GSM8K is numeric-MCQ rather than standard exact-match, and turbo3 is a KV-cache runtime policy rather than a standalone GGUF weight file type.

The BF16 GGUF itself was also re-evaluated headlessly on the same 128-case / 8-fold CV gate. K=bf16,V=turbo3 and K=q8_0,V=turbo3 tie for the top mean at 0.5547 +/- 0.0662, so the selected BF16 serving policy is K=bf16,V=turbo3. The four measured groups have Friedman p=0.943806 and pairwise p-values >= 0.750973; this is a best-evidence BF16 selection, not a statistically dominant winner. K=q8_0,V=turbo4 needed a longer single-policy startup retry and then completed. Artifacts are under eval/best_bf16_*, including eval/best_bf16_selection.md and eval/best_bf16_lm_eval_gguf_kv_cv_report.md.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "zapabobouj/qwen3.5-elt-l3"
tok = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, torch_dtype="auto", device_map="auto")

For GGUF downloads:

hf download zapabobouj/qwen3.5-elt-l3 elt-lm-qwen35-side-stem-aha-ilsd-l3-Q8_0.gguf

Provenance

• Source method: Elastic Looped Transformer / ILSD workflow.
• Base model: huihui-ai/Huihui-Qwen3.5-4B-Claude-4.6-Opus-abliterated.
• Export date: 2026-05-17.
• README statistics: see eval/l3_readme_stats.json and eval/l3_readme_stats.md.