GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
mahidwasif
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
mahidwasif
Model Tree
Qwen/Qwen2.5-3B-Instruct
Input Modalities
Output Modalities
Supported Functionality
This project fine-tunes a LoRA adapter for Qwen2.5-3B-Instruct to evaluate technical interview answers against a provided question and scoring rubric.
The model is designed to return a JSON object with:
score: integer from 0 to 5max_score: integer, always 5justification: explanation for the scorestrengths: demonstrated strengthsimprovements: suggested improvementsQwen/Qwen2.5-3B-Instruct
The same held-out test examples were used for the base and fine-tuned model. Results below should be interpreted as agreement with synthetic expected labels, not as proof of real-world interviewer reliability.
| Metric | Base model | Fine-tuned model |
|---|---|---|
| Valid JSON (%) | 98.33% | 100.0% |
| Schema-valid (%) | 98.33% | 100.0% |
| Score MAE | 1.8814 | 2.0333 |
| Exact score accuracy (%) | 15.25% | 21.67% |
Research, experimentation, and educational demonstrations of small-model fine-tuning for structured technical-answer evaluation.
Load the compatible base model and adapter using a compatible Transformers and PEFT setup. The adapter is not a standalone full model.
Check the base model's license and applicable terms before redistributing or using this adapter. This model card does not grant rights beyond those terms.