ministral-3-3b-cn-gguf
Quantized GGUF release for COVAS:NEXT fine-tuning experiments on unsloth/Ministral-3-3B-Instruct-2512.
Files
model-q8_0-v0.8.gguf: Q8_0 GGUF exported from the latest mask-boundary-fixed run.
Source run
- Training output:
output/ministral_newmix_maskfix_e4_es_bs1
- Dataset:
data/training_stitched_full.jsonl
- Key fix: train-time label masking boundary now starts after the later of
[/INST] and [/TOOL_RESULTS].
Eval snapshot (strict-6, 47 cases, 282 candidate responses)
- Overall:
244/282 (86.5%)
- Tool calling:
90/114 (78.9%)
- Event reaction:
37/42 (88.1%)
- Contextual QA:
59/60 (98.3%)
- Tool result summarization:
58/66 (87.9%)
Tool-calling micro metrics:
- Call made:
114/114 (100.0%)
- Name correct:
102/114 (89.5%)
- Args correct:
102/114 (89.5%)
See experiment details in docs/experiments.md.
