DrStrangel0ve/adaptfm-qwen35-single-path-artifacts

AdaptFM Qwen3.5 Single-Path Closeout - 2026-06-15

This folder contains the final locally preserved package artifacts from the single-serving AdaptFM/Qwen3.5 work. The GCP VM and its persistent disk were deleted after these files were copied and hash-verified.

Public Hugging Face mirror:

https://huggingface.co/DrStrangel0ve/adaptfm-qwen35-single-path-artifacts

Preserved Tarballs

Fast-risk staged candidate

• File: image-stock-awq-marlin-spec12-prefix-seq16-allocnofree-v2-thinkcap1024.tar.gz
• Image: gcp2-single-stock-awq-marlin-aliasnorm-v2-spec12-prefix-seq16-allocnofree-v2-thinkcap1024:latest
• Size: 14,512,183,483 bytes
• SHA256: ffa4a939cdbeb925faea414381aa3a9568e0aa700b1893d3236c3561c90b1737
• Official upload staging id: 01bef639
• Staged S3 key: submissions/AFM-7f3gf5c9/01bef639/image.tar.gz
• Official state at shutdown: UPLOAD_COMPLETE; /submit retries returned HTTP 400.
• Local evidence: fast local MMLU/IFEval gates passed, but extended synthetic GPQA-shape stress was risky: 28/32 under 60s, 4/32 timeout/over-60, 2/32 missing final answer, 26/32 synthetic letter correctness.
• Verdict: preserve for research/provenance only. This was the final fast-risk candidate, not a quality-safe candidate.

Validity-first fallback

• File: image-stock-awq-marlin-v1-nospec-thinkcap2048.tar.gz
• Image: gcp2-single-stock-awq-marlin-aliasnorm-v1-nospec-thinkcap2048:latest
• Size: 14,512,172,065 bytes
• SHA256: 9235f746473f76a95d7fab4d1c312c4d9a968e63c08e876ebd7c8a1b9d0d3a1d
• Official upload staging id: 9deae075
• Staged S3 key: submissions/AFM-7f3gf5c9/9deae075/image.tar.gz
• Official state: upload completed but /submit was daily-limit blocked.
• Local evidence: MMLU-Pro 2% 0.7323, IFEval 2% 0.8421, GPQA-shape synthetic 8/8 under 60s; latency roughly 2961 / 5941 / 6146 ms.
• Verdict: best preserved validity fallback, but too slow for the speed target.

Cloud Shutdown

• Original GPU VM: instance-20260531-095309
• Project: project-b8d90673-56c0-46ed-a97
• Zone: us-central1-a
• Temporary extractor VM: qwen-artifact-extract-20260615
• Final state: both instances deleted; original 500 GB persistent disk deleted.

Single-Path Work Summary

The single-path phase started after the rule clarification that split serving would be filtered or banned. Historical dual-path submissions reached about 20x speedup with good quality, but used separate latency and quality paths and were no longer eligible.

After the reset, every promoted candidate used one model artifact and one serving path. Endpoint compatibility work focused on making /ping, /invocations, /v1/completions, and /v1/chat/completions robust without routing to another model. Major fixes included blank authorization stripping, alias normalization, chat/completion payload remapping, buffering behavior for non-stream requests, and preserving thinking flags for GPQA-style streaming.

The BF16/hardened line fixed endpoint failures but was too slow or blank-quality under official conditions. The quality-LoRA no-spec line produced a visible official result with MMLU and IFEval passing, but GPQA failed badly. That made GPQA thinking behavior the main quality blocker.

The AWQ/Marlin W4A16 line was the main speed path for A10G compatibility. It looked strong in small local gates, but official full-harness results exposed severe quality collapse for speculative variants:

• 99ecdd25 / spec14 no-prefix: official MMLU collapsed near zero.
• b9ec849e / spec14 prefix + thinking cap: official 6.535x, but MMLU 0.007, IFEval 0.809, GPQA 0.131.

The final lesson is that narrow internal subsets were not predictive enough for the official full A10G harness. The safest artifact was no-spec AWQ with a thinking cap, but it sacrificed speed. The fastest legal artifacts remained quality-risky because speculative decoding and allocator/prefix-cache behavior could preserve local subset quality while corrupting full official MMLU/GPQA.

No official benchmark data was used for training, calibration, distillation, prompt tuning, routing, or cached answers. Official and benchmark-like use was measurement-only.