GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
gokul1019
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
gokul1019
Model Tree
unsloth/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
unsloth/Qwen2.5-1.5B-Instruct (apache-2.0)Evaluated against 8 golden questions (never trained on), judged for correctness/completeness against reference notes by an independent LLM judge, compared to the stock (non-finetuned) base model and to the original RAG-teacher system.
| Arm | Correctness | Completeness |
|---|---|---|
| Base (stock unsloth/Qwen2.5-1.5B-Instruct) | 0/8 | 0/8 |
| This finetuned model | 0/8 | 0/8 |
| RAG-teacher (real retrieval + citations) | 0.5-0.75 (varied across real runs) | -- |
The finetuned model did not measurably beat the stock base model. Both scored 0/8 on a strict binary correctness judge, across every real eval run, including after scaling the training set from 191 to 331 real pairs and retraining from scratch -- the result did not change, and the specific failure pattern (confidently claiming coverage of an out-of-scope question) persisted identically across both adapters.
Two plausible causes were tested and ruled out with real controlled experiments, not assumed:
The qualitative difference that did survive finetuning: this model stayed topically relevant to aviation on every question, while the stock base model regularly hallucinated unrelated content (e.g. confusing aviation V-speed terminology with cloud-computing/virtualization terms). Finetuning changed the model's behavior (answer directly, stay in-domain) without teaching it precise factual recall -- consistent with LoRA being sample-efficient at behavior transfer but not at encoding large volumes of new factual knowledge into a small (1.5B parameter) model without retrieval.
The real, achievable ceiling for this exact judge/question set isn't 100% either -- the RAG-teacher itself, with real retrieval and citations, only scored 0.5-0.75 correct across real runs (generation isn't fully deterministic even at judge temperature 0.0).
Full per-question answers, methodology, and the diagnostic investigation: eval report · project history · learning notes.
Do not use this model as a source of real aviation facts. The measured result above shows it does not reliably produce correct answers to aviation knowledge questions. This is a portfolio/research artifact demonstrating a distillation pipeline and its measured limits, not a production aviation reference. For real FAA Airplane Flying Handbook questions, use the original RAG system linked above, which is grounded in real, citable source material.
apache-2.0, matching the base model (unsloth/Qwen2.5-1.5B-Instruct). This is a
finetuned derivative; no additional restrictions are added.