Quat3rnion/Signal-3.8-Flash-Next-HGN-v2-abliterated

Signal + original Orca BF16 writers149 — native HGN v2 baseline + delta

This is requantization from an already prequantized Signal Q8_0 GGUF, not conversion from original Signal BF16. Decoding and freshly encoding prequantized values cannot recover the original full-precision Signal weights and can introduce additional quantization error. Native v2, not HT43, is used.

No broad model-quality or behavioral-transfer claim is made. This publication verifies conversion fidelity, exact file identities and reconstructability, not refusal removal, safety, reasoning, long-context quality, broad runtime compatibility or equivalence to an upstream checkpoint. Separately operated local runtime and benchmark work is not a portable compatibility or model-quality guarantee.

This repository is a differential release, not a standalone model directory. Four byte-identical tokenizer files must be obtained from the immutable Peonist baseline below. Signal-specific template/configuration and all three distinct HGN files remain here. Historical queue and first-upload authorizations are included for provenance only and do not resume any retired work; the current differential policy is recorded in provenance/differential-publication.json.

Pinned inputs and auxiliary provenance

• Original Signal input: agentionai/Signal-3.8-Flash-Next-GGUF @ 4d30467d9f087cd88640762808c05d14131a329b, Q8_0/Signal-3.8-Flash-Next-Q8_0.gguf, 188,225,032,736 bytes, SHA-256 98fb2414cdffccc16b9ff4222aae60015b1284cdea19a15232f389cae1a66cdd. The whole-file acquisition receipt and execution-time source hash gate are included in provenance.
• Original base: Qwen/Qwen3.8-Flash-Next @ de4b8e4d43b917e7706784d8bb445c9af86a3540. Initial MTP and vision inputs are extracted from this exact original base BF16 source, not a stock HGN or another derivative. The pinned receipt lists upstream hashes/blob identities, extracted tensor identities and exact artifact hashes. These auxiliary weights are not claimed to be Signal-trained or measured-equivalent to Signal.
• Extracted original base MTP input: 31 tensors, 5,214,306,816 bytes, SHA-256 8b440f527c1eea97e7e11b90aa869513307e70679d5f00688260a72b664e5b32. Extracted original base vision input: 333 tensors, 897,916,416 bytes, SHA-256 cf721a42a9e50747d93bd39e3a9e6f24f33ac6599fe17b9f5c843fb7e3b2203e. These are input provenance hashes, not hashes of the newly encoded HGN outputs.
• Main MTP is embedded in model.hgn according to the verified plan (31 encoded MTP tensors); there is no separately substituted stock MTP file. Selected donor replacements, when applicable, are applied before encoding this main-file MTP.
• The n-gram table is supplied by the table embedded in the original Signal Q8 input, then freshly encoded into the generated n-gram HGN.
• The approved Signal/base asset package retains the actual Signal GGUF chat template; the original base template is different and is not silently substituted. The reconstructed tokenizer/chat_template.jinja and tokenizer/configuration/processor assets are byte-identical to the approved inputs. Signal/base vocabulary coordinates were checked, not whole tokenizer JSON equality, shared training lineage or behavior.
• Exact Signal preprocessing equivalence is not established: GGUF pre label does not expose full regex; exact preprocessing identity cannot be established from retained metadata. Assets retain verified original Qwen tokenizer preprocessing and exact token coordinates/merge order. Retaining the actual template and matched vocabulary/merge order does not reconstruct an unavailable complete Signal tokenizer preprocessing pipeline.

Original Orca BF16 donor composition

• Donor: orcarouter/Qwen3.8-Flash-Next-Uncensored @ 8336e613ea508b13c2159bd0f68965d97a606b95.
• Exactly 149 canonical tensors, enumerated in provenance/donor-writers149.txt, are replaced with original donor BF16 values before a fresh conversion from the original Signal GGUF plus the approved auxiliary inputs. Selector SHA-256: edae870c34e575ca07e00a389e647884993f05bab73295fa25d2607785549e2c.
• This does not convert the already generated plain HGN. Applicable selected MTP tensors also receive the donor values before encoding; unselected inputs retain the original Signal/base-auxiliary composition.
• Signal/base tokenizer and other approved target assets are retained. The checks establish vocabulary-coordinate compatibility, not identical tokenizer JSON, common training lineage, preserved behavior or model quality. Donor tokenizer assets are not installed.
• The historical Swift-associated selector and producer-hash references provide only names and pinned original donor hashes. No Swift tensor values, tokenizer/assets or license obligations from weight inclusion are introduced. The historical reference SHA-256 is 813cdc880785b54478944ce3ec31b108c779fac9acaa477b2fc56d78230d7f85; its old transformation procedure is not applied.
• 'abliterated' identifies the donor's origin only. Refusal removal and all behavioral transfer are untested. This is experimental direct tensor replacement, not a refusal-direction projection or an Orca-minus-base delta.

Conversion and full artifact verification

The verified conversion service signal-v2-donor149-hip-20261014T203725Z.service exited successfully. Conversion and full verify --decode-backend hip both returned exit code 0. Options were:

--profile v2 --ht-backend hip --i4r-backend hip --q4cp-backend cpu --seed 0 --threads 2 --ht-beam 128

Manifest schema 2, provenance attested, 1,658 source tensors, 2,349 output tensors, and 118,785,466,780 payload bytes across the complete output set were verified. Converter executable SHA-256: bb64ff5dcc6e9ab4800c095cedec538aba1bfd08a456cfc8cb67278a2aa4c807. Manifest SHA-256: b4065f67bb4b222a45a63c52ea7d1157309285c477fb0ed5cc25d645d0f9d872. Full HIP verification report SHA-256: 3b3da43e5455316ed56f53f8a8fa3109d13874ae7f6ea0eebcd07248e7360773.

Every reconstructed conversion file retains its original relative filename. Ten runtime files, including the distinct main/n-gram/vision HGN files, Signal template/configuration/processor assets and conversion-manifest.json, are published here; four identical tokenizer files are supplied by the pinned baseline. All three HGN files were freshly encoded; no stock HGN tensor payloads, additive weight delta, refusal projection, calibration Hessians or HT43 artifacts are reused. This is a file-level differential package, not a binary patch or tensor delta. No weight-data copies are made for publication staging.

The raw conversion completion receipt, verification report and process arguments, plan, approved input gate, exact output inventory, auxiliary/acquisition receipts, build/qualification receipts and publication approval/configuration are under provenance/. They describe the actual conversion, not historical model benchmarks. Local paths in these artifacts are provenance references, not portable runtime instructions.

Immutable baseline, exact omissions and reconstruction

Baseline: **peonist-ai/halogen-qwen3.8-flash-next @ 5cc17cea1a10b502b4a5db41d8d2b1a0f22dba26. This is the documented stock native v2 revision, not the vendor's mutable main or an HT43 checkpoint. It is a distribution source for four identical tokenizer files, separate from the original BF16 conversion inputs above.

Exactly 4 files / 22,885,540 bytes (21.83 MiB) are omitted from this repository:

• tokenizer/generation_config.json — 202 bytes, SHA-256 e70c136c1b78ddc1fb0905bac8e733a4dc448d4f852a5dd75143fffc70be550e.
• tokenizer/merges.txt — 3,353,259 bytes, SHA-256 a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d.
• tokenizer/tokenizer.json — 12,809,320 bytes, SHA-256 0997f410c57a1f4e53b09e4be8f4a172d90edd9564368fb0847030937229b9f3.
• tokenizer/vocab.json — 6,722,759 bytes, SHA-256 ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003.

Every original payload was compared against the pinned baseline's size/SHA-256 identities; ordinary baseline files were downloaded at that revision and SHA-256 hashed. All three large HGN files differ: the main and n-gram differ in size/hash, and vision differs in hash despite equal size. They cannot be omitted or substituted, so this preference saves only the tokenizer bytes above, not tens of gigabytes. Required license texts, NOTICE and conversion provenance are retained regardless of duplication policy.

assembly-manifest.json lists all 14 reconstructed runtime files, their exact destinations/sizes/SHA-256 and whether each comes from this release or the pinned baseline. provenance/differential-publication.json includes the complete original 41-file comparison and baseline identities. Do not substitute the baseline chat template, tokenizer configuration, main, n-gram, vision or separate MTP file: this main already embeds its own verified 31-tensor MTP head.

Assemble a usable, hash-checked model directory

Install huggingface_hub with the hf CLI and Python 3 on Linux. Set REV to this release's verified 40-hex Hub commit, not main; the assembler rejects non-immutable revisions. Download only the small instructions/legal/provenance files first:

REV=REPLACE_WITH_VERIFIED_40_HEX_RELEASE_COMMIT
hf download Quat3rnion/Signal-3.8-Flash-Next-HGN-v2-abliterated --revision "$REV" \
  --include README.md assemble.py assembly-manifest.json 'LICENSE*' NOTICE 'provenance/*' \
  --local-dir ./signal-release
python ./signal-release/assemble.py --manifest ./signal-release/assembly-manifest.json \
  --revision "$REV" --output-dir ./Signal-3.8-Flash-Next-HGN-v2-abliterated
cp ./signal-release/README.md ./signal-release/assemble.py ./signal-release/assembly-manifest.json \
  ./signal-release/LICENSE* ./signal-release/NOTICE ./Signal-3.8-Flash-Next-HGN-v2-abliterated/
cp -a ./signal-release/provenance ./Signal-3.8-Flash-Next-HGN-v2-abliterated/

The assembler downloads each manifest source at its immutable commit through the official Hub SDK, verifies every complete file's SHA-256 and size, and atomically installs a new directory. It uses hardlinks only from downloaded SDK cache files on the same filesystem to avoid a second large payload copy; the assembled directory therefore contains regular files, not external symlinks. Keep the output on the cache's filesystem, or explicitly add --copy to permit a cross-filesystem copy. Existing output directories are never overwritten. Do not point the normal download/cache mode at protected local conversion outputs; the separate --verify-local SOURCE_MAP_JSON mode only opens/hashes actual files and creates no links or payload copies.

Keep the license texts, NOTICE, manifest and provenance with any redistributed assembled directory. Use a compatible Halogen native-v2 engine and the assembled Signal tokenizer/template. This is not a Transformers checkpoint; these assembly instructions do not claim a portable engine deployment recipe or broad runtime qualification.

Complete reconstructed runtime integrity

Reconstructed destination	Exact bytes	SHA-256	Immutable source
config.json	4,745	889658f2508e8c61d409b02e70e0d78d8d4452ec65aaafbe129805d213d2e74b	published:config.json
conversion-manifest.json	4,039,251	b4065f67bb4b222a45a63c52ea7d1157309285c477fb0ed5cc25d645d0f9d872	published:conversion-manifest.json
model.hgn	66,687,683,040	4365a7bcd2d2232046cb40b3121b33ed71e7d27d7517227794603d9888ce716f	published:model.hgn
preprocessor_config.json	390	27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516	published:preprocessor_config.json
qwen38-flash-next-ngram.hgn	51,200,246,084	93604e21a5fb2046a20b29ed321b4e3842f69316970387534a7c346fbf3e51ac	published:qwen38-flash-next-ngram.hgn
qwen38-flash-next-vision.hgn	897,916,416	b6ca6716eaab601773f7155701bb09f39c72922caa21620bab577f0281f2ea2a	published:qwen38-flash-next-vision.hgn
tokenizer/chat_template.jinja	9,993	12827f24b742ea4e80cdc12dbcf9622227056b9f797252a3149263d4f9aaadce	published:tokenizer/chat_template.jinja
tokenizer/config.json	4,745	889658f2508e8c61d409b02e70e0d78d8d4452ec65aaafbe129805d213d2e74b	published:tokenizer/config.json
tokenizer/generation_config.json	202	e70c136c1b78ddc1fb0905bac8e733a4dc448d4f852a5dd75143fffc70be550e	baseline:tokenizer/generation_config.json
tokenizer/merges.txt	3,353,259	a9d356d7bdf1ef4949e3e748e95b8e10ad9d4e2e838eddc38a0a7b6b94d1db8d	baseline:tokenizer/merges.txt
tokenizer/tokenizer.json	12,809,320	0997f410c57a1f4e53b09e4be8f4a172d90edd9564368fb0847030937229b9f3	baseline:tokenizer/tokenizer.json
tokenizer/tokenizer_config.json	17,418	eb7404e61aa4a747b838093da7ee2289c94a866279a9f99f8e28edd6ed3280fc	published:tokenizer/tokenizer_config.json
tokenizer/vocab.json	6,722,759	ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003	baseline:tokenizer/vocab.json
video_preprocessor_config.json	385	7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13	published:video_preprocessor_config.json

The table covers all reconstructed runtime outputs. Durable publication receipts separately verify every actually published file, including this card, the assembly tools, legal/provenance records and unchanged Hub-generated .gitattributes. They also recheck the omitted files against the immutable baseline, the local actual-path full-file assembly hash smoke and adjacent parent-guarded commit ancestry. Preupload activity alone is not a successful publication. Historical source-conversion inventories describe the complete local output, not a claim that all those bytes are duplicated in this differential repository.

Licenses and attribution

This package is not blanket Apache-2.0 licensed. Exact source license texts and NOTICE are included; derivative processing does not remove their obligations:

• LICENSE: exact original Qwen/Qwen3.8-Flash-Next Qwen Community License 1.0.

• LICENSE-SIGNAL: exact pinned agentionai/Signal-3.8-Flash-Next-GGUF Qwen Community License 1.0 repository license.

• LICENSE-ORCA: exact original orcarouter/Qwen3.8-Flash-Next-Uncensored Qwen Community License 1.0 text, not an Apache-license inference from metadata.

Read the full texts, including attribution conditions and the requirement for a separate Qwen license for commercial Model-as-a-Service / AI Work Assistant business use. The summary is not legal advice. Source names identify provenance and do not imply endorsement. The Halogen engine has separate terms; no portable deployment configuration or broad engine compatibility is claimed here. This is an unofficial experimental artifact. Users are responsible for applicable licenses/laws and appropriate safety and abuse-prevention controls.