Day21-Qwen3.5-4B-LoRA-TruongVietAnh API & Inference Endpoint

Evaluation

Fresh Kaggle evaluation (not the historical Colab run):

Table with columns: Run, Target field accuracy, Regression keyword recall, JSON format, Latency ms
Run	Target field accuracy	Regression keyword recall	JSON format	Latency ms
(a) base + naive prompt	0.0	0.7578	0.0	2956.8
(b) base + optimized prompt	0.765	0.7578	1.0	939.8
(c) LoRA fine-tune	0.97	0.5889	1.0	1577.7
Verdict: FAIL. Reasons: {"passed": false, "reasons": ["general capability regressed by 0.169 (tolerance 0.020). See deck §6.3 — add 1-5% replay data."], "target_delta": 0.20499999999999996, "regression_delta": -0.16888888888888887}

Limitations and loading

One seed on a small synthetic corpus; not a general-purpose or production-validated model. Inspect regression and formatting, not training loss alone. Load the base model with Transformers, then use PeftModel.from_pretrained(base, this repository). Load the tokenizer from unsloth/Qwen3.5-4B and use its chat template. T4 requires fp16 rather than bf16. Reproduction source: https://github.com/vietanh2005-tva/Day21-Track3-Finetuning-Lab-TruongVietAnh

Evaluation

Fresh Kaggle evaluation (not the historical Colab run):

Table with columns: Run, Target field accuracy, Regression keyword recall, JSON format, Latency ms
Run	Target field accuracy	Regression keyword recall	JSON format	Latency ms
(a) base + naive prompt	0.0	0.7578	0.0	2956.8
(b) base + optimized prompt	0.765	0.7578	1.0	939.8
(c) LoRA fine-tune	0.97	0.5889	1.0	1577.7
Verdict: FAIL. Reasons: {"passed": false, "reasons": ["general capability regressed by 0.169 (tolerance 0.020). See deck §6.3 — add 1-5% replay data."], "target_delta": 0.20499999999999996, "regression_delta": -0.16888888888888887}

Day21-Qwen3.5-4B-LoRA-TruongVietAnh

README

Evaluation

Limitations and loading

Explore FriendliAI today

README

Evaluation

Limitations and loading