Huihui-Qwen3.5-2B-abliterated-PALW
This is the PALW-converted form of an existing model for the MISAKA chain: one PALWART3 .palwart artifact (an index file plus segment files). It is not a newly trained or fine-tuned model and makes no claim about the model's quality or safety. The source model is an uncensored variant (refusal behaviour reduced upstream).
Scope: text generation only
Token ids in, next-token logits out. Vision towers and projectors (mmproj), image/audio/video embedders and processors, and multi-token-prediction heads are out of scope: not computed, not covered, not admitted as claims, not rewarded. The artifact refuses them by name (check_text_scope_v3). Out of scope in this checkpoint: vision tower model.visual.* (ignored by the adapter); MTP head mtp.* (ignored by the adapter).
Provenance
| Base model | Qwen/Qwen3.5-2B — Qwen team (Alibaba Cloud) |
| Derivation | abliteration — huihui-ai |
| Source repository | huihui-ai/Huihui-Qwen3.5-2B-abliterated |
| Source revision (pinned) | b2e291a65f29a9b148981fa5299caea5d35bd4c8 |
| Architecture | Qwen3_5ForConditionalGeneration (qwen35-gdn-hybrid) |
| License | apache-2.0 |
Source weight files (publisher LFS SHA-256, checked before conversion):
model.safetensors-00001-of-00001.safetensors— 4,548,221,488 bytes —568f9529262f887dfab700a2b7de1f9607bbf3d88a797748351ed6480fe3d24c
The artifact
.palwartroot (identity):136b823b66c0fb9b248dd9d733ceabf511164c91503b3fea7e2cd5f878d7c21ccefb0a3cb19eb02d543e137d720d7b53817121d2155ae04b346eed2122399a0a- index
Huihui-Qwen3.5-2B-abliterated-PALW.palwartSHA-2567241eaf1dcd2421718d08d238f4c986c55c504ce95bba85349cccc47a97d01d9 - weights section: tir-inventory, inner root (3)
a644954878ae42dda2cbcc9de879ccabb6e52db99d96f74d8dda57cedc4b4db411d8d3978fdd211f3fb0c42ac1c1f7235a3d27e3299980b2051295ecf7a18c2e - segments: 2, 2,544,144,327 bytes
Huihui-Qwen3.5-2B-abliterated-PALW.00000.palwseg— 2,147,483,648 bytes — SHA-256cc5266c81ef4cb04a398b906ecf7499f3f6d78d940f30b24bd1cc14e51c83fe1Huihui-Qwen3.5-2B-abliterated-PALW.00001.palwseg— 396,660,679 bytes — SHA-256f208ee43a08a52dab99670d0f0881a6bd72aca4028024de85433286aac6c8f97
PALWART3 holds a version header, a canonical manifest (manifest.json here is a readable copy), the program record (TIR program and kernel id, or the A16 / PALWQ361 engine), the per-architecture weight section, the tokenizer, the chat template, the source configuration, the license texts and the calibration inputs. Legacy formats (.palwart A16, .palwtir, .palwq36) are carried unmodified as sections and can be re-extracted byte for byte.
Check it (no node needed): palwart verify <stem>.palwart (segments beside the index), palwart extract-legacy.
Conversion
- converter:
tir; options:{"calib": {"chunks": 1, "docs": ["docs/rfc/0002-palw-tensor-ir.md"], "kind": "recurrent", "len": 32768}, "context": 32768, "gdn_core_wide": true} - served context: 32,768 tokens
- calibration: 1 sequence(s) of 32768 tokens from
docs/rfc/0002-palw-tensor-ir.md(git blob7bb6e09f61b2) atd52d6c56ff33, tokenised byhuihui-ai/Huihui-Qwen3.5-2B-abliterated@b2e291a65f29tokenizer.json; token file SHA-2564aa2f715a18140fc2bb007db57c3ad7f20f23ad4feff78da636dabcf41ea900f; the token file and the measured statistics are carried in the artifact - A Gated DeltaNet class is calibrated on one sequence as long as the context it serves (freeze-v1 §5.2): here one 32,768-token sequence; a longer context is outside this class
- fit against the float reference (held-out text, 128 positions): mean KL 0.00267, top-1 0.977 (tolerance: KL <= 0.1, top-1 >= 0.8)
- long-context fit (held-out 32,768 positions, every one compared): mean KL 0.00566, top-1 0.958, worst window mean KL 0.00793, late/early 1.9007658595599624
Verification on MISAKA
End-to-end adjudicable: claims over this model are checked by sampled deterministic re-execution, localized with public evidence and decided by the deterministic court. MISAKA never uses cryptographic proofs of model computation (PRINCIPLES §8). Hashes and Merkle roots here authenticate bytes only.
License and attribution
Apache License 2.0 — the full text is in LICENSE. huihui-ai/Huihui-Qwen3.5-2B-abliterated@b2e291a65f29a9b148981fa5299caea5d35bd4c8 (abliteration by huihui-ai) of Qwen/Qwen3.5-2B by Qwen team (Alibaba Cloud).