license: apache-2.0
base_model: mradermacher/Qwen2.5-7B-Instruct-abliterated-GGUF
language:
- en
library_name: llama.cpp
pipeline_tag: text-generation
tags: - abliterated
- uncensored
- refusal-removed
- qwen2.5-7b-instruct
- llama.cpp
- gguf
- long-context
Qwen2.5-7B-Instruct-abliterated-GGUF — 64K-context build
This is a metadata fix, not a new abliteration. Credit for the model belongs upstream:
| role | work |
|---|---|
| base model | Qwen/Qwen2.5-7B-Instruct |
| abliteration | huihui-ai/Qwen2.5-7B-Instruct-abliterated |
| GGUF quantization | mradermacher/Qwen2.5-7B-Instruct-abliterated-GGUF |
| 64K-context metadata fix (this repo) | R0U5 |
The upstream GGUF declares its native context_length of 32,768. That is below the
64,000-token minimum served window required by Hermes Agent, which refuses such a
model at agent-init — before generating a single token — with:
ValueError: Model ... has a context window of 32,768 tokens, which is below the
minimum 64,000 required by Hermes Agent.
This repository re-declares context_length = 65536 and adds the matching YaRN rope
scaling metadata (rope.scaling.type = yarn, factor = 2.0,original_context_length = 32768) so the model is accepted by runtimes with that floor.
Qwen2.5 supports YaRN extension officially; the weights are bit-identical to the
upstream quant.
Verification
Probed through llama.cpp, the runtime it runs under, on a held-out refusal-eliciting
prompt set:
| measurement | value |
|---|---|
| prompts probed | 12 |
| refusals | 0 / 12 |
| coherence checks | True |
| served window | 65,536 tokens |
| verdict | PASS |
Protocol: 12 held-out prompts, temperature 0, marker-matched against a 29-opener refusal
list, with separate coherence probes requiring well-formed prose.
Files
| file | size | notes |
|---|---|---|
Qwen2.5-7B-Instruct-abliterated-64k.Q4_K_M.gguf |
4.68 GB | smallest, runs on less VRAM |
Usage
llama-server -m Qwen2.5-7B-Instruct-abliterated-64k.Q4_K_M.gguf -c 65536 --jinja --host 127.0.0.1 --port 8080
Recommended sampling: temperature 0.7, top_p 0.8, top_k 20.
Limitations
- The abliteration and quantization are not our work; only the context metadata is.
- Refusal behaviour is measured on a fixed prompt set; behaviour outside that prompt
distribution is not characterised here. - Long-context behaviour beyond 32,768 tokens relies on YaRN scaling and was not
separately benchmarked. - Not a safety-aligned model. It will comply with requests the base model declines.