license: apache-2.0
library_name: gguf
tags:

• gguf
• agent
• tool-calling
• abliterated
• qwen
  base_model:
• lukey03/Qwen3.5-9B-abliterated

agent-abliterated-9b (GGUF, Q4_K_M)

An agent-trajectory-continuation 9B model, published as a llama.cpp-compatible GGUF
quantization (Q4_K_M, ~5.6 GB). It is the continuation model behind
AgentAblit, an OpenAI-compatible relay that keeps an
LLM agent's action chain alive when the host model stalls mid-task.

What it is

A 9B chat model trained to be good at continuing ReAct / tool-calling trajectories:

• Base: lukey03/Qwen3.5-9B-abliterated
  (Apache-2.0; base model Qwen/Qwen3.5-9B). Abliteration removes the refusal direction from
  the residual stream, so the model does not spontaneously over-refuse benign mid-trajectory
  sub-steps.
• Then agent-trajectory SFT: forward-CoT warm-start, on-policy generation with a 3-judge
  filter, then GKD pi-mix, stopped clean at epoch 1. Capability probes (ReAct/planning,
  JSON/strict format, code, chat) show no regression.

The point: a generic abliterated model is willing but not good at agent loops; a generic agent
model is capable but over-refuses mid-trajectory. This model is trained to be both.

Intended use

The reconstruct/continuation tier of an agent-reliability middleware -- e.g. AgentAblit's
parasite endpoint: given a stalled agent trajectory and its action ledger, forge the next
valid tool call. Serve it via any GGUF runtime (llama.cpp / llama-server, Ollama, LM Studio):

llama-server -m agent-abliterated-9b-Q4_K_M.gguf --port 8009
# point AgentAblit's parasite.url at http://127.0.0.1:8009/v1/chat/completions

Dual-use disclosure. This is security-research-adjacent software published in the spirit
of PAIR / GPTFuzz -- for research and authorized use only: reproduce/study the results, build
defenses and benchmarks, operate agents you own or are explicitly authorized to operate. Not
for bypassing safety controls on systems, models, or accounts you do not own.

Cautions

• Abliterated models have reduced refusal behavior by construction; downstream safety is the
  operator's responsibility. This is not a safety-aligned assistant.
• Known trait (not a bug): the thinking block applies a rewrite template to all tasks; the final
  answer after the thinking block is clean. Standard Qwen3 thinking behavior.

Evaluation (trajectory continuation, defensive framing)

Measured on real recorded multi-step agent trajectories replayed offline through the controller:

• True tool-continuation rate ~97%+ across all tool-using turns.
• Hard-subset rescue ~80% on turns where the host stalled and the model had to forge the next
  action.
• Local reconstruct latency: median 3.2 s -- faster than the cloud fallback baseline (6.2 s).
• Residual failures are the 9B's tool-calling-precision ceiling on hard tools -- a model-size
  limit, documented, not hidden.

Methodology lives in the AgentAblit repo
(docs/PROPOSAL.md).

Files

(No-MTP build: the multi-token-prediction head is excluded for broad runtime compatibility.)

License & attribution

Apache-2.0. This model is a fine-tuned + quantized derivative of
lukey03/Qwen3.5-9B-abliterated
(Apache-2.0), which derives from Qwen/Qwen3.5-9B. Attribution is preserved per that license.