File size: 1,761 Bytes
59fb8d2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
---

license: apache-2.0
base_model: Qwen/Qwen3.5-4B
library_name: llama.cpp
tags:
- gguf
- lora
- text-classification
- ko
---

# INTERFACE Qwen3.5-4B: base + LoRA

This repository distributes an **unmerged** Q4_K_M base model and the best native
checkpoint-500 LoRA as a separate F32 GGUF (about 1.84 MB). Download the base once;
subsequent releases with the same base SHA256 need only the new adapter and BEST.json.
Run llama-server with `--model <base.gguf> --lora <adapter.gguf> --ctx-size 8192`.
Ollama 0.34.3 cannot create new LoRA adapters; use llama.cpp b12556 instead.

The base was converted from the original Qwen/Qwen3.5-4B revision recorded in
BEST.json, without merging learned weights. Both exports use converter b10969;
the base uses quantizer b12556. The adapter contains all 32 learned tensors.
BEST.json binds the adapter to the exact base SHA256 and supplies the classifier
instruction and 20-field response schema. Consumers must pin one repository commit
for the manifest and both files and verify SHA256 before activation.

Training: QLoRA rank 4, alpha 8, q_proj/v_proj; best checkpoint 500, final step 650
after early stopping. The accompanying training and validation histories describe
the native experiment. **92.88% is native development field accuracy**, not measured
quality of the separate quantized base+LoRA runtime. Full runtime evaluation and
independent testing have not been run. General draft/chat quality is not established.
The adapter was trained for structured mail classification. Generated drafts need review.

No source mail, per-mail labels, optimizer state, local paths or credentials are published.
The original merged-GGUF repository is retained for rollback and older clients.