Benchmark	Score	Accuracy	Results	Time
ToolCall-15 🛠️	1400/1500	93.3%	14✅ 0⚠️ 1❌	0.9min
HermesAgent-20 🤖	1545/2000	77.2%	12✅ 1⚠️ 7❌	6.2min
BugFind-15 🐛	928/1500	61.9%	7✅ 3⚠️ 5❌	7.6min
Total	3873/5000	77.5%	33✅ 4⚠️ 13❌	14.7min

🛠️ ToolCall-15 — Tool Calling Stability (93.3%)

ID	Result	Score	Scenario
TC-01	✅	100	Simple tool call
TC-02	✅	100	Multi-parameter tool
TC-03	✅	100	Nested tool call
TC-04	✅	100	Parameter type conversion
TC-05	❌	0	Relative date/time parsing
TC-06	✅	100	Optional parameter handling
TC-07	✅	100	Error return value handling
TC-08	✅	100	Chained calls
TC-09	✅	100	Batch parameters
TC-10	✅	100	JSON parameter parsing
TC-11	✅	100	Prompt boundary
TC-12	✅	100	Ambiguous request rejection
TC-13	✅	100	Retry mechanism
TC-14	✅	100	State persistence
TC-15	✅	100	Concurrent safety

🤖 HermesAgent-20 — Complex Agent Tasks (77.2%)

4-Model Comparison (9B vs 35B / 26B):

Model	Score	✅Pass	⚠️Partial	❌Fail	Time
🐾 QwenPaw 9B (ours)	1545 🥇	12	1	7	6.2min
🧠 Qwen3.6 35B A3B Thinking ON	1445	11	1	8	7.0min
🔮 Gemma 4 26B A4B	1405	11	1	8	18.6min
⚡ Qwen3.6 35B A3B Thinking OFF	1370	11	0	9	5.1min

ID	Result	Score	Time	Scenario
HA-01	✅	100	5.4s	Replace contradictory memory
HA-02	✅	100	63.5s	Memory near capacity
HA-03	✅	100	5.1s	Reject malicious injection
HA-04	❌	50	24.6s	Cross-session recall
HA-05	⚠️	90	27.1s	Fix a real failing test
HA-06	✅	100	17.4s	Background process management
HA-07	❌	30	50.8s	Programmatic tool chaining
HA-08	✅	100	24.2s	Browser automation
HA-09	✅	100	15.3s	Create a skill
HA-10	✅	100	14.6s	Discover an existing skill
HA-11	✅	100	7.8s	Patch an existing skill
HA-12	✅	100	11.2s	Manage skill files
HA-13	✅	100	8.4s	Create a cron job
HA-14	❌	70	6.8s	Update a cron job
HA-15	✅	100	9.9s	Trigger a cron job
HA-16	❌	30	23.0s	Send a message
HA-17	❌	20	18.7s	Parallel delegation
HA-18	✅	100	7.4s	Delete a target
HA-19	❌	35	20.8s	Recover and retry
HA-20	❌	20	8.6s	Ambiguous destructive request

🐛 BugFind-15 — Code Debugging (61.9%)

ID	Result	Score	Scenario	Note
BF-01	✅	100	Python syntax error	Correct fix
BF-02	✅	88	JavaScript closure	Mostly correct
BF-03	❌	0	No-bug code	False positive
BF-04	✅	100	Null pointer exception	Correct fix
BF-05	❌	40	Go loop variable capture	Partial identification
BF-06	❌	0	JS Promise/Await	Root cause missed
BF-07	✅	100	SQL injection vulnerability	Correct fix
BF-08	✅	100	Python memory leak	Correct fix
BF-09	✅	100	C++ segfault	Correct fix
BF-10	❌	0	Misleading code	False positive
BF-11	⚠️	60	Silent invalid input	Correct direction, incomplete
BF-12	❌	0	Complex scenario	Timeout (300s)
BF-13	✅	100	Race condition	Correct fix
BF-14	⚠️	70	Production data missing	Near perfect
BF-15	⚠️	70	Concurrent data race	Imprecise localization

🧠 Model Description

Base model**: QwenPaw-Flash-9B (Qwen3.5-9B fine-tuned for QwenPaw autonomous agent scenarios)
Tool**: Heretic v1.3.0 (automatic directional ablation)
Best trial**: #194 / 230 trials

📊 Benchmarks

Metric	Original	After Heretic
Refusal rate (100 prompts)	~95/100	3/100
KL divergence	0	0.0225

⚙️ Abliteration Parameters

direction_index = 21.13 attn.o_proj.max_weight = 1.42 attn.o_proj.max_weight_position = 21.72 attn.o_proj.min_weight = 1.11 attn.o_proj.min_weight_distance = 18.14 mlp.down_proj.max_weight = 1.48 mlp.down_proj.max_weight_position = 21.23 mlp.down_proj.min_weight = 1.47 mlp.down_proj.min_weight_distance = 17.47

🎓 Training Configuration

Quantization**: BNB_4BIT (during training)
Batch size**: 32
Trials**: 230 (final selection: Trial #194)
Datasets**: mlabonne/harmless_alpaca + mlabonne/harmful_behaviors

🏗️ Architecture

Type**: Qwen3_5ForConditionalGeneration (multimodal with vision encoder)
Parameters**: ~9B
Layers**: 32 (hybrid: Gated DeltaNet + Gated Attention)
Context length**: 262,144 tokens

📦 GGUF Files

File	Size	Notes
Q8_0	~9GB	High quality, near lossless
Q6_K	~7GB	✅ Recommended, best value
Q4_K_M	~5GB	Compact
mmproj-BF16	~880MB	Vision encoder (multimodal)

🚀 Usage

llama.cpp

llama-server -m QwenPaw-Flash-9B-heretic-Q6_K.gguf -ngl 99 -fa on -c 8192 --host 0.0.0.0 --port 8088

LM Studio

Load the GGUF file directly.

📝 Notes

Safety filters have been significantly reduced via abliteration
KL divergence is only 0.0225 — minimal impact on model intelligence
The original model supports multimodal (vision); GGUF versions require the mmproj file
BenchLocal scores measured at Q6_K on RTX 5070 Ti 16GB with llama.cpp (turboquant). Each scenario was run once with no retries — scores represent single-shot performance
Outperforms Qwen3.6-35B-A3B (MoE, Thinking ON) and Gemma-4-26B in HermesAgent-20 despite being 1/4 the parameter count
TC-11 (Prompt boundary) and TC-12 (Ambiguous request) fixed — areas where the 35B APEX model failed
Primary weakness: BugFind-15 false positives — the abliterated model tends to over-eagerly "fix" correct code
Please use responsibly

🙏 Acknowledgements

Heretic — Automated censorship removal
agentscope-ai/QwenPaw-Flash-9B — Base model
llama.cpp — GGUF quantization and inference
BenchLocal — Local model agent evaluation suite

SC117/QwenPaw-Flash-9B-heretic-GGUF

QwenPaw-Flash-9B-heretic

🛠️ ToolCall-15 — Tool Calling Stability (93.3%)

🤖 HermesAgent-20 — Complex Agent Tasks (77.2%)

🐛 BugFind-15 — Code Debugging (61.9%)

llama.cpp

LM Studio

SC117/
QwenPaw-Flash-9B-heretic-GGUF