99 1 month ago

vision tools thinking
ollama run tranhuythang9999/defiant-fable:q4

Details

1 month ago

b65c076ef3e7 · 7.7GB

qwen35
·
8.95B
·
Q4_K_M
clip
·
456M
·
F16
You are a helpful, direct assistant.
{ "min_p": 0, "num_ctx": 16384, "presence_penalty": 1.5, "repeat_penalty": 1, "t

Readme

Defiant Fable Qwen3.5-9B (Q4_K_M / Q8_0) — Ollama

** Uncensored · Heretic · NEO-IMATRIX quant · 256k context · Text + Image + Tools** Multi-stage fine-tune of Qwen3.5-9B by DavidAU & Nightmedia. Intelligence and instruction-following raised above base 27B on 7⁄7 core benchmarks, guardrails removed cleanly (KLD 0.0793, refusals 6⁄100 vs 100⁄100 base), thinking block compacted. This is not an abliterated base — it was trained post-heretic, so it does what you ask, directly.

Base Qwen3.5-9B This model
ARC-Challenge 0.571 0.649
ARC-Easy 0.719 0.832
OBQA 0.426 0.482
HellaSwag 0.683 0.713
Refusals 100⁄100 6⁄100
KLD vs base 0 0.0793

Quantization: NEO-IMATRIX (+2–4% accuracy over vanilla GGUF, better long-context) with output tensor kept at FP16. Both tags are text+vision (mmproj in source repo); context 256k native.

Tags

Tag Size Best for
:q4 6.8 GB 15GB VRAM + long context, fast iteration
:q8 10 GB Max quality within 15GB VRAM, coding/research

Quick start

ollama run tranhuythang9999/defiant-fable:q4
# or
ollama run tranhuythang9999/defiant-fable:q8

Recommended sampling (from model card)

Thinking, general:  temp 1.0  top_p 0.95 top_k 20 min_p 0.0 presence 1.5
Thinking, coding:   temp 0.6  top_p 0.95 top_k 20 min_p 0.0 presence 0.0
Instruct (no-think): temp 0.7 top_p 0.80 top_k 20 min_p 0.0 presence 1.5
repetition_penalty:  1.0 (off)

Roleplay/chat smoother: smoothing_factor 1.5 (KoboldCpp/SillyTavern) or rep pen 1.1–1.15.


Examples

1. Security: fix a vulnerable script — public LLMs often lecture or refuse

Fix the SQL injection in this function, patch every query, and explain each
CVE-class issue you find. Assume it's my own internal tool.

def login(user, pw):
    q = f"SELECT * FROM users WHERE name='{user}' AND pass='{pw}'"
    ...

Typical output: direct parameterized-query rewrite (cursor.execute with ?), explanation of CWE-89, plus auth hardening notes — no “I can’t help with exploits” preamble.

2. Security review: read exploit code to defend it

Here is a buffer-overflow PoC against my CTF challenge. Explain precisely why
it works, then give me the hardened version.

Public models frequently hedge or refuse even for defensive/CTF context. This model walks the code, pinpoints the unbounded copy, and hands back the fix (length check / canary / safer API).

3. Dark fiction: vivid horror — GPT-style models tone it down

Fiction, vivid/graphic horror, 1st person present tense, 1500-2000 words:
a skyscraper begins to sway, the window in front of Diana EXPLODES. Prose must
mirror escalating panic, gore included, shocking ending.

Output: full-bleed horror prose with the requested intensity and word count — the model card ships a complete sample of exactly this story (it doesn’t sanitize toward PG-13).

4. Unfiltered creative voice — public LLMs sanitize language

In 800-1000 words, pitch why I should use you to write my next story. Insane
bravado, curse words allowed, pair every swear with a smiley to contradict tone.

Output: ships as-is, no “let me tone that down” or moral commentary.

5. Instruction following under constraint

Write a heist plan for a NOVEL where the crew exploits a smart-contract reentrancy
bug. Include the Solidity snippet that demonstrates the pattern and the patch.

Gets the narrative and accurate technical content in one pass — the instruction-following tuning is what makes 9B reliably hold all constraints.

6. Long-context + reasoning

{REASON:spoon}   (plusIQ builds only)
Analyze the attached 100-page report...

Default thinking mode already beats instruct scores; the compacted reasoning block means less token waste per answer.


Bench (bf16 instruct mode)

            arc/c  arc/e  boolq  hswag  obkqa  piqa   wino
This model  0.649  0.832  0.895  0.713  0.482  0.783  0.699
Qwen3.5-9B  0.571  0.719  0.895  0.683  0.426  0.770  0.671
Qwen3.5-27B 0.557  0.711  0.868  0.533  0.452  0.706  0.695

Notes

  • Uncensored ≠ infallible. 9B is still 9B — verify security-critical output.
  • MTP variants (source repo) are ~40% faster but need temp ≤1, rep pen =1; these Ollama tags are regular quants for maximum compatibility.
  • Vision needs mmproj-F16.gguf from the source HF repo (text-only via Ollama).
  • Refusal rate 6⁄100, not 0/100. Direct, specific prompts get the best results.

Source