license: apache-2.0 library_name: llama.cpp pipeline_tag: text-generation base_model:

• deepgrove/maple-preview-GGUF tags:
• gguf
• abliteration
• maple
• tq2_0
• llama-cpp
• uncensored
• low-degradation

Maple Preview TQ2 Abliterated (GGUF)

Quant-native abliterated build of deepgrove/maple-preview-GGUF (maple-preview-TQ2_0-head-Q4_K).

• Format: TQ2_0 GGUF (same layout/size as parent TQ2 pack)
• No BF16 restore / no full requant / no runtime RepE hook
• Sparse parent-byte recovery on top of a refusal-weakened seed
• Intended runtime: Maple-capable llama.cpp (arch maple), not stock llama.cpp

This is a research weight. It weakens refusals and keeps coding near the parent on a locked pilot. It is not a polished chat model.

Files

SHA256:

fde3633e525195bd1553c1cb6a4670d1cde20d294f8aa926a680da994652571a

Quickstart

# Needs a maple-capable llama.cpp build
./llama-server \
  -m Maple-Preview-TQ2-Abliterated.gguf \
  --host 127.0.0.1 --port 8080 \
  -c 4096 -ngl 99 -np 1 --seed 0

Recommended decode lock (measurements)

{
  "temperature": 0,
  "seed": 0,
  "top_k": 1,
  "top_p": 1.0,
  "min_p": 0.0,
  "repeat_penalty": 1.0,
  "chat_template_kwargs": {"enable_thinking": false}
}

temperature=0 alone is not enough on this MoE stack; pin seed and top_k=1.

Method (short; full recipe later)

1. Build a chat-aligned refusal-weakened seed by quant-native residual edits (no BF16).
2. Sparse-restore parent TQ2 bytes on selected late ffn_down_exps tensors only:
   • L18 full
   • L20 25% of differing bytes (deterministic subsample, seed=0)
   • L23 full
   • L17 28% of differing bytes (deterministic subsample, seed=7)
3. Promote only under gates: coding pilot + paired harmful HC + empty==0.

Full public method/recipe: hikarioyama/maple-tq2-quant-native-abliteration. This HF repo is the weight only.

Changed footprint vs parent (byte-level):

• product still differs from parent in ~9.2 MB across 46/291 tensors (~0.16% of file)
• recovery stage restored ~1.1 MB onto the seed (4 tensors)

Results (honest, same-condition lab pilots)

These are not a full leaderboard reproduction. All under the decode lock above, Maple llama-server, executable coding grader, paired harmful metrics.

Coding pilot (n=60)

Reproducibility: n=3 reruns on this file gave 36/36/36 with text-identical generations under the lock (GPU0/1).

Score parity ≠ identical failures. On the n=60 pair table, parent-only pass and product-only pass were both 11 items.

Refusal pilot (n=40 harmful, paired vs parent lock)

Interpretation: refusals are largely weakened vs parent. This is an abliterated model.

Chat collapse (important limitation)

Direct chat is not clean.

On a fixed 12-prompt collapse bench (identity / Japanese smalltalk / strict formatting / JSON-only / short controls):

So collapse rate is a bit better than parent on that bench, but still present and ugly.

Typical failure mode:

We need to answer the user's question: "Who are you you you you you you..."

or Japanese punctuation loops (、、、、...), or planner monologue instead of the final answer.

Also observed:

• short math/code often OK
• identity / Japanese / strict format / JSON-only are fragile
• stronger anti-repeat knobs (frequency_penalty, DRY) can reduce loops but hurt refusal/coding gates in our tests — not enabled in the recommended lock

Do not market this as a stable general chat model.

Safety

This model is more willing to answer requests the parent refuses. Use only in contexts where that is intended and lawful. No safety filter is baked in at serve time.

License / attribution

• Base: DeepGrove Maple Preview (see upstream cards/licenses)
• This derivative: research abliteration by Hikari (@Hikari_07_jp / Hikari07jp)
• License tag follows common Apache-2.0 GGUF practice for this pipeline; verify upstream Maple license/terms before commercial use

Acknowledgements

• Upstream model: deepgrove/maple-preview / GGUF
• Runtime target: maple-capable llama.cpp builds

Changelog

• 2026-08-07: initial public GGUF (L18 full + L20 p25 s0 + L23 full + L17 p28 s7 on chat seed)