license: apache-2.0 base_model:

• SC117/occamy-1.0-abliterated-FIT-GGUF
• Accio-Lab/occamy-1.0 base_model_relation: quantized pipeline_tag: text-generation library_name: gguf language:
• en
• zh tags:
• gguf
• ICE
• ice-quant
• imatrix
• sangreal
• moe
• qwen35moe
• vision
• abliterated
• uncensored
• occamy

Occamy-1.0 Abliterated — Sangreal ICE

Four ICE 1.6 tiers of Accio-Lab's Occamy-1.0 (35B total, 3B active, Qwen3.5-MoE architecture), abliterated trunk by SC117, quantized against Sangreal — a purpose-built 77-bucket calibration corpus — with an importance matrix measured on this abliterated Occamy itself, and shipping Qwen-Sharp v22.5.0-arrow.3.1 — our agent-tuned chat template, a heavily modified derivative of peculiar-ragdoll's Qwen-Sharp (spec below). Vision works through Occamy's own projector, mmproj-Occamy-1.0-Q8_0.gguf.

⚠️ Uncensored. This is an abliterated checkpoint — refusal behaviour has been suppressed in the weights. Sandbox it at the OS level and control its network and code-execution access; with no refusal backstop, a prompt injection from a hostile page or third-party code has nothing to stop it.

The trunk

Accio-Lab/occamy-1.0 is Accio-Lab's agentic co-work fine-tune of Qwen3.6-35B-A3B: hybrid gated-delta-net + full attention, 256 routed experts top-8 beside a shared expert on every layer, a native vision tower, and a 262K-token context. Every tier here is quantized from SC117's abliterated BF16 of it, occamy-1.0-abliterated-BF16.gguf in SC117/occamy-1.0-abliterated-FIT-GGUF — per their card, an input-gated ablation that applies per-expert deltas.

At a glance — against SC117 at the same size

Mean KLD to BF16, one session. Size-matched = SC117's own files for this same BF16 (five FIT tiers and APEX-I-Quality), interpolated (log-linear) at each ICE file's exact byte size; negative = ICE closer to BF16. Paired = the same 64 chunks scored for both files against the same reference.

Mean KLD vs file size, code at 16K context: Sangreal ICE 1.6 against SC117's FIT and APEX-I files

Code at 16K context (code.test.raw, n_ctx 16384 × 8), mean KLD to BF16 vs file size; lower-left is better. Raw numbers in Table 2 below and in measurements/2026-10-05/.

Measurements

One binary (llama.cpp 3653e6d), one reference (SC117's BF16, the file every row was quantized from), one session (2026-10-05, 4× RTX PRO 6000 Blackwell 96 GB). 64 chunks at n_ctx 2048; the long-context table uses 8 chunks at n_ctx 16384. Every file below was measured in this session. Raw logs: measurements/2026-10-05/.

Table 1 — Code (code.test.raw)

PPL(base) 2.180124 · BF16 = 100

Table 2 — Code at long context (code.test.raw, n_ctx 16384 × 8)

PPL(base) 2.262030 · BF16 = 100

Compare rows within this table only.

Table 3 — English text (WikiText-2)

PPL(base) 6.442774 · BF16 = 100

overall = 0.70/(1 + meanKLD) + 0.30 * sameTop1, ×100. BF16 = 100. The same composite as the Ornith-1.5 ICE card: 70% on how close the whole output distribution stays, 30% on agreement about the argmax.

Read the tail columns as shape, not order. 99.9% KLD is roughly the 33rd-worst token of 32,768 — an extreme order statistic with large sampling variance, so it inverts between adjacent files without meaning anything. Rank on mean KLD.

Two bpw columns. Only 8 of 256 experts fire per token, so a bit in ffn_*_exps is worth ~3% of a bit in attention, the shared expert or the output head. active bpw weights by that (token embedding excluded: it is a lookup, not a matmul); file bpw is just size ÷ parameters.

Don't compare the tables to each other. Code is more predictable text, so every file scores about half the divergence on it. Compare rows within a table.

Every row is measured against SC117's BF16 — the file their tiers and ours were both cut from. Rivals are SC117's own quants of the same weights, so a difference here is recipe and calibration, nothing else. These numbers are not comparable to SC117's card, which reports a five-domain macro KL at n_ctx 512. SC117's file names count GiB (FIT-QUALITY-17G = 17.50 GiB); every size on this card is decimal GB.

What Sangreal is

The calibration corpus these tiers were quantized against. Built from primary sources, not assembled from eaddario's set, bartowski's calibration_datav5, or any other ready-made calibration file. 77 buckets, 9,533 documents, 47,443,549 tokens, sha256 85a6b823a6762bf6….

The full render is 92,663 chunks. These tiers use a 10,000-chunk pass over it, measured on this abliterated Occamy itself - 20x the 500-chunk imatrix behind SC117's files.

Split: technical 70.2 · science+domain 20.1 · language 9.7.

Synthetic is 21.7% by bytes and sits in exactly three buckets — teacher:k2-horizon, teacher:deepseek-v4-pro, teacher:hermes. Everything outside reasoning and agentic is real text.

Language — 11 languages, 5 non-Latin scripts

The spread across the whole block is 0.06 points, and that is deliberate: language's job here is activation, not ranking. Each bucket sits just above the saturation floor at 44.9% consumption, so these buckets select for the first time — the legacy v1 spec had consumed 96% of turkish and 92% of english.

Mathematics — absent from the pool until this cycle

Sourced from the LaTeX algebraic geometry, mathlib4, UniMath, math-comp, open-web-math, AutoMathText and proof-pile.

The imatrix and the recipes are not published with this release.

What's inside

ICE = Isolation of Compounding Error: allocate bits by how far a quantization error travels, not by how large the activations are. Routers stay F32, SSM decay gates stay F32, the KV-cached projections stay F16, the always-on dense path stays Q8_0, and the budget goes to the routed expert bank — most of the parameters, but only 8-of-256 active per token. ICE 1.6 keeps token_embd at Q8_0 (its error propagates with context) and moves the output head to Q6_K (its error stays flat from 2K to 16K). Method and the cases where it does not win: gbuzhf/ICE-quantization.

No MTP head. Occamy-1.0 ships none (mtp_num_hidden_layers: 0), and neither does SC117's BF16. These tiers are the 40-block trunk; there is no built-in draft model.

Chat template: Qwen-Sharp v22.5.0-arrow.3.1 (ours; spec below). Occamy-1.0 keeps Qwen3.6's chat format — its own template is byte-identical to Qwen3.6-35B-A3B's — and Qwen3.6's tokenizer, so arrow.3.1 applies as-is: the reply opener in every thinking mode, the tool-call format and the tool-result wrapping match what Occamy was trained on, checked mode by mode. Thinking is kept across turns by default, which is what Occamy's card recommends for agent runs. The delimiters are native tokens — checked on every tier at signing.

Vision. Occamy's vision tower is byte-identical to Qwen3.6-35B-A3B's, and the abliteration did not touch it. mmproj-Occamy-1.0-Q8_0.gguf is that tower at Q8_0 (334 tensors). Load it with --mmproj mmproj-Occamy-1.0-Q8_0.gguf.

Chat template — Qwen-Sharp v22.5.0-arrow.3.1

Lineage: Qwen official → froggeric/Qwen-Fixed-Chat-Templates (thinking retention, reasoning effort, XML/JSON tool formats, error coaching) → peculiar-ragdoll/Qwen-Sharp v22.5.0 (terse block) → arrow (ours). Both upstreams are Apache-2.0; what arrow changes is stated here.

• Plain chat is untouched: with no tools, arrow.3.1 renders byte-identical to Sharp v22.5.0.
• Agent mode (tools present): a decisive-agent directive — reuse what you already have, locate before reading, fix in place — plus fixed error coaching and a skills line naming the real skill tools.
• Arrow notes: at most one short note per tool result, chosen by priority — identical repeated call, full rewrite of an already-known file, hard-won fix (same tool failed twice, then passed), step back (3rd / 6th error), stall (12 calls without a change), plan (at 25 calls), and a .arrow memory nudge every 10 calls. Every note is computed from earlier messages only, so the prompt cache stays valid. ~700 prompt tokens in agent mode.
• Switches (chat_template_kwargs): decisive: false drops the agent directive; arrow_memory: false turns the notes off, "always" forces them without file tools. Sharp's terse and froggeric's reasoning_effort / thinking controls are unchanged.

Files

• measurements/2026-10-05/ — raw llama-perplexity output for every row of every table (4 ICE tiers + 6 SC117 files × 3 columns), the three BF16 base runs, and the session summary.

Credits

KLD measures fidelity to this repo's reference and nothing else — not reasoning, tool use or speed. It is also within-model: every file here is measured against SC117's BF16, so these values are not comparable to another repo's. Compare across models with PPL.