linkd-dsl-qwen3-4b

Merged full-weights release (Qwen/Qwen3-4B-Instruct-2507 + the linkd-search DSL LoRA, ericmao/linkd-dsl-qwen3-4b-lora). Turns free-form people-search queries into MongoDB find filters for the Berkeley.profilematch collection.

Held-out eval (200 queries, result-grounded gpt-oss-120b judge): reward 0.1507 vs gpt-5.5-medium 0.1043, 99% valid DSL, 0.76 s p50 generation on an RTX 4090.

Usage

vllm serve ericmao/linkd-dsl-qwen3-4b \
  --max-model-len 2048 \
  --speculative-config '{"method":"ngram","num_speculative_tokens":8,"prompt_lookup_max":4,"prompt_lookup_min":2}'

Call with the exact production system prompt (linkd-search repo, slm/common.py:SYSTEM_PROMPT), temperature=0; the completion is a raw JSON Mongo filter for collection.find(filter).limit(20).