FastLane LoRA adapters (Qwen2.5-3B-Instruct)

LoRA adapters for the small model of FastLane, a voice agent that books, changes and cancels movie tickets, bus tickets and restaurant tables for Indian-English callers. The small model answers every caller turn first and passes only hard turns to a large model or to staff. Code, data generator and evaluation: https://github.com/Sankalp0109/FastLane-Adaptive-Voice-Agent

Folder Training data Used for
v3/ all three domains, data v3 FastLane in the final runs
v5/ all three domains, data v5 (round-trip speech noise) FastLane v5, matched pair for A8
mb-v5/ movies and buses only, data v5 ablation A8 (new domain: restaurants)

Training: QLoRA on one Kaggle T4, 4-bit base, rank 16, alpha 32, dropout 0.05, all seven projection matrices. The adapters expect FastLane's own prompt and JSON output format; they are not general chat models.

Use

hf download Sankalp0109/fastlane-adapters --local-dir adapters
fastlane train finish adapters/v3 --name fastlane-qwen2.5-3b      # 4-bit MLX model in runs/
fastlane serve --model runs/mlx-fastlane-qwen2.5-3b

LMA Group Project 2026, team Chaturvedi and Sons (Noel Alex Jacob, Aviral Tyagi, Sankalp Chaturvedi, KSPSVLN Siddardha Kumar Kavuri), IIIT Hyderabad.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Sankalp0109/fastlane-adapters

Base model

Qwen/Qwen2.5-3B
Adapter
(1,514)
this model