Downloads · 30 days
0
lab21-qwen35-4b-lora-correct
dkvippro2k5/lab21-qwen35-4b-lora-correct
lab21-qwen35-4b-lora-correct is a machine learning model from dkvippro2k5. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This adapter was produced by a training run whose finalloss diverged to ~3,059,379 instead of a healthy sub-1.0 value. The resulting model generates only repeated ! characters at inference time — it does not produce v…
Access
Public
Updated Aug 22, 2026
Repo size
130 MB
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors130 MB · 100%
At a glance
- Access
- Public
- Created
- Aug 22, 2026
- Updated
- Aug 22, 2026
- SHA
- ff9d9b05
Model card
From the Hugging Face model README
Lab 21 — LoRA adapter correct (Qwen3.5-4B) — ⚠ KNOWN BROKEN
This adapter was produced by a training run whose final_loss diverged to
~3,059,379 instead of a healthy sub-1.0 value. The resulting model generates
only repeated ! characters at inference time — it does not produce
valid JSON and should not be used for the intended ticket-triage task.
Root cause (suspected): the training script's loss_type was patched
from "chunked_nll" to "nll" to work around a trl/Kaggle compatibility
bug (AttributeError: 'functools.partial' object has no attribute '__func__' in _patch_chunked_ce_lm_head). The patched loss path appears
to not normalize by the supervised token count, causing the loss/gradient
scale to explode.
This adapter is uploaded here only as supporting evidence for the
accompanying lab report (submission/REPORT.md in the GitHub repo below),
not as a usable model artifact.
- Base model:
unsloth/Qwen3.5-4B - Config:
all-linear, r=16, LR 1e-4, fp16, 30 steps - Full report / results: https://github.com/dkvippro2k5/Day21-Track3-Finetuning-Lab-2A202601724-DuongVanKien
- Created
- Aug 22, 2026
- Updated
- Aug 22, 2026