GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
lakshmianne
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
lakshmianne
Model Tree
Qwen/Qwen3-0.6B
Input Modalities
Output Modalities
Supported Functionality
Use this repository to reproduce research on domain adaptation versus capability retention, compare CGAF with matched LoRA/rehearsal/hard-projection baselines, or extend the gate with an explicit retention constraint.
Do not treat the current method as state of the art or deploy it as a safety-preservation mechanism.
CGAF computes domain and anchor gradients for trainable adapter groups. Negative alignment triggers a smooth, temperature-controlled projection. The confirmed configuration uses layer grouping, temperature 0.1, and minimum conflict 0.0.
Across Qwen3-0.6B QLoRA runs with seeds 42, 123, and 456:
| Method | Domain F1 | Anchor F1 | Harmonic F1 |
|---|---|---|---|
| Rehearsal | 0.2803 | 0.2152 | 0.2432 |
| LoRA | 0.2725 | 0.2175 | 0.2418 |
| CGAF | 0.2813 | 0.2094 | 0.2399 |
| Hard projection | 0.2752 | 0.2128 | 0.2396 |
CGAF improved domain adaptation but did not improve the balanced primary metric. All paired seed-bootstrap intervals against LoRA included zero.
The final experiment uses license-governed examples from databricks/databricks-dolly-15k: summarization for domain adaptation and closed/general QA as retention anchors. Exact split construction is recorded in the experiment reports.
One small model, one English dataset, three seeds, and small final-test splits limit generalization. Token F1 is incomplete. Current Hugging Face artifacts do not include adapter weights; reproduce them from the code and configuration.
QLoRA reduces memory requirements, but gradient-conflict methods require two backward paths and are slower than plain LoRA. Early stopping reduced unnecessary updates.
Until an archival paper identifier exists, cite the software using CITATION.cff and link the GitHub repository.
This repository contains the validation-selected seed-42 CGAF adapter. Release evaluation: domain F1 0.2553, anchor F1 0.2025, harmonic F1 0.2259.