DavidBShan

linkd-search-qwen3.6-35b-a3b-grpo100-lora

Deploy Dedicated

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Get help setting up a custom Dedicated Endpoints.

Talk with our engineer to get a quote for reserved GPU instances with discounts.

README

License: apache-2.0

linkd-search GRPO LoRA — grpo_000100 (lineage best)

Rank-64 LoRA adapter (alpha 32, all-linear) over Qwen/Qwen3.6-35B-A3B, trained with the Freesolo SFT -> GRPO pipeline on the linkd-search task: translating natural-language people-search requests into MongoDB find filters for the Berkeley.profilematch collection.

Tinker checkpoint: tinker://4173b059-c42e-53ef-8782-293086486655:train:0/sampler_weights/grpo_000100
Lineage-best final_eval score: 0.6672 ("Executable Search Quality", 347 rows, real query execution + LLM judge)
Renderer at train/eval time: qwen3_5 (thinking enabled; only the final non-thinking text is scored)
Generation settings used for evaluation: temperature 0.0, max_tokens 2048

Model provider

DavidBShan

Model tree

Base

Qwen/Qwen3.6-35B-A3B

Adapter

this model

Modalities

Input

Video, Text, Image

Output

Text