Qwen2.5-1.5B Ward Distilled v4 Balanced
Final deployable model
- GGUF:
/root/autodl-tmp/qwen/models/Qwen2.5-1.5B-Ward-Distilled-v4-GGUF/qwen2.5-1.5b-ward-v4-q6_k.gguf
- Quantization:
Q6_K
- File size:
1,272,739,456 bytes (1.273 GB, 1213.78 MiB)
- SHA-256:
f7ffc9751378a8f67bed3a6a8872f0f20c40b48cc4223b5560e225bd547ca8a0
- SHA-256 short form:
f7ffc9751378a8f6
The previous v3 model remains available and was not overwritten.
What changed
- Built a leakage-free 5,000-row mixed training set (
v4.1.0):
- 2,000 ward-domain rows
- 1,500 official GSM8K training rows
- 1,500 MBPP-derived code execution/output reasoning rows
- Trained a general-capability LoRA from the original Qwen2.5-1.5B-Instruct base.
- Fused the established ward LoRA and the new general LoRA by exact rank concatenation.
- Screened multiple fusion weights. The selected balance is:
- ward delta weight:
0.50
- general delta weight:
0.75
- Merged the selected adapter and converted it to Q6_K GGUF.
The frozen 600-question benchmark was not used as training data. The builder
reported zero exact overlap with the frozen GSM8K test questions, and CRUXEval
was not used as a training source.
Full evaluation results
All general-capability figures below use the same frozen 600-question benchmark,
temperature 0, and the same symmetric relaxed answer extractor.
| Model | GSM8K (200) | CRUXEval-O (200) | BBH (200) | Macro |
|---|
| Original Qwen2.5-1.5B | 69.0% | 26.0% | 31.0% | 42.0% |
| Previous ward v3 Q6_K | 47.5% | 16.5% | 41.0% | 35.0% |
| New ward v4 Q6_K | 60.0% | 25.0% | 35.0% | 40.0% |
New v4 versus original 1.5B retention:
- Math:
86.96%
- Code:
96.15%
- Natural-language/logic:
112.90%
- Macro retention:
95.24%
New v4 versus previous v3 Q6_K:
- Math:
+12.5 percentage points
- Code:
+8.5 percentage points
- BBH:
-6.0 percentage points
- Macro:
+5.0 percentage points
Ward-domain full evaluation (200 samples):
- Cloud judgment accuracy:
99.0%
- Cloud judgment macro-F1:
99.14%
- Event urgency accuracy:
82.0%
- Format valid rate:
100.0%
- Mean reference similarity:
0.4673
This is a trade-off model: math and code improved substantially over v3, while
BBH and ward urgency are slightly lower than v3. The core ward judgment remains
99% accurate on this test set.
Memory and speed check
Strict single-instance deployment profile:
1--n-gpu-layers 20 --ctx-size 4096 --parallel 1 \
2--cache-type-k q8_0 --cache-type-v q8_0
- Idle GPU memory:
1438 MiB
- GPU memory after one request:
1442 MiB
- One short ward request total time:
0.370 s
- Prompt processing:
89 ms
- Generation: about
44.4 tokens/s
The test establishes runtime GPU memory below 1.5 GiB for this exact profile.
It does not establish a 75% TTFT reduction because a paired original-model TTFT
test has not yet been run.
Start and use
1qwen-start-v4
2qwen-status
3qwen-chat 'B07患者疑似跌倒，置信度0.93，网络断开，请给出简洁处置建议。'
Stop the service with:
OpenAI-compatible endpoint: http://127.0.0.1:8000/v1.
Evidence files
- General Q6 full report:
datasets/ward-nlu-500-v1/distillation/v4/general-v4-q6-full.json
- General Q6 raw samples:
datasets/ward-nlu-500-v1/distillation/v4/general-v4-q6-full.samples.jsonl
- Ward Q6 full report:
datasets/ward-nlu-500-v1/distillation/v4/domain-v4-q6-full.json
- Selected pre-quantization full report:
datasets/ward-nlu-500-v1/distillation/v4/general-combo-a050-full.json
- Selected pre-quantization ward report:
datasets/ward-nlu-500-v1/distillation/v4/domain-combo-a050-full.json
- Mixed-data manifest:
datasets/ward-nlu-500-v1/distillation/v4/manifest-v4.1.0.json
- Adapter fusion manifest:
/root/autodl-tmp/qwen/models/Qwen2.5-1.5B-Ward-LoRA-v4-combo-a050/combination_manifest.json
