Downloads · 30 days
1K
83% of all-time downloads
xu2409324124/qwen3-4b-latex-correction-v3-gguf-q4-k-m
qwen3-4b-latex-correction-v3-gguf-q4-k-m is a text generation model from xu2409324124. Use it when you need the model to write or continue text. The card lists the license as apache-2.0.
The historical repository name contains v3; its main now serves v5.2 checkpoint-30. The previous v3 GGUF remains at v3-archive-20261023, commit 7350f8000b05bee313b95b93ccae45ac3add7383.
Downloads · 30 days
1K
83% of all-time downloads
All-time downloads
1.2K
Public
Repo size
5 GB
Likes
0
Public
Click a slice to open those files.
.gguf2.5 GB · 100%
From the Hugging Face model README
The historical repository name contains v3; its main now serves v5.2 checkpoint-30. The previous v3 GGUF remains at v3-archive-20261023, commit 7350f8000b05bee313b95b93ccae45ac3add7383.
The full merged BF16 weights were converted to GGUF and quantized to Q4_K_M using llama.cpp 063d9c156e816ae3cf62db01f429a07a099afe97. The GGUF embeds the tokenizer and Qwen chat template. This is a full model, not an adapter. It has passed CPU file loading and tokenization validation; quantized task accuracy has not been measured.
Download qwen3-4b-latex-correction-v5.2-q4_k_m.gguf and Modelfile into one folder, then:
ollama create qwen-latex-v5.2 -f Modelfile
ollama run qwen-latex-v5.2
The Modelfile supplies the versioned formula-repair system prompt, num_predict 256, greedy decoding, and a 4096-token context. It relies on the Qwen template embedded in GGUF. Use the OpenAI-compatible Ollama endpoint with your own selected model name when integrating a CLI. No remote service or automatic background startup is included.
For custom clients, send system-prompt.txt as the system message. Preserve input symbols and verify results independently. Long generation limits can amplify repetition.
cdbee75f17c01a7cc42f958dc650907174af0554.xu2409324124/qwen3-4b-latex-correction-v5-2, revision 3721b25f40f3e9383d99b2991ba9dd0e4749dc86, path runs/v52-20261023T073851Z-d239b61a/sft/checkpoint-30.Reviewed formula-correction development accuracy: base 33/56 (58.93%) → checkpoint-30 48/56 (85.71%); equal-weight formula-group mean 77.22% → 92.12%. Development contains only four formula groups. The final test set has not been run. The second epoch did not improve development accuracy.
General mathematics results are mixed: GSM8K first 200 at 4K output budget 90.0% → 89.5%; MATH-500 first 200 at 4K conservative answer-validation score 56.5% → 66.0% (format-sensitive); AIME25 all 30 at 32K 40.0% → 36.67%. SFT hit the 32K output limit on 9/30 versus base 3/30; extensive boxed-answer repetition occurred on 8 versus 3. This does not establish overfitting or universal reasoning improvement. MATH 16K and the proposed three-item 8K rerun were not executed.