finder-1-qwen3.8-27b

finder-1 (Qwen3.8-27B, RL post-trained for people search)

finder-1 is Qwen/Qwen3.8-27B post-trained with LoRA GRPO to act as a web-search agent that finds real people for B2B lead generation and candidate sourcing and returns structured, evidenced profiles (name, title, company, location, LinkedIn URL, publicly listed email, bio, verified extra facts, per-criterion evidence quotes with source URLs).

This checkpoint is the merged bf16 weights of iteration RL5. It is meant to be run with the accompanying agent harness: the system prompt (system_prompt.md), the tool schemas (tools_schema.py: web_search, fetch, python, places, scholar, company_registry, finish) and a token-exact tool loop with a per-query tool budget (12 calls; 16 for open-ended list requests). Serving was done with vLLM (--reasoning-parser qwen3 --tool-call-parser qwen3_coder, reasoning effort "low").

Results: LessieAI/people-search-bench (B2B + recruiting + deterministic)

Scored with the unmodified bench harness (criteria-grounded verification, Tavily advanced search), judge gpt-5-mini, on the fixed stratified 30-query subset (10 per category). Bench queries were never used for training.

Per category (official harness): deterministic 68.1, B2B 74.7, recruiting 72.6; task completion 100%.

Caveats: 30-query subset, one official scoring pass, gpt-5-mini judge instead of the harness default Gemini. Judge nondeterminism alone moves the 30-query Overall by about +-2 points between scorings of identical submissions, and the verifier's search backend shifts it by 2-3 points, so differences under ~3 points are not meaningful. Most of the score comes from the harness and output contract: the base model with the same harness scores 69.6; RL added ~1-2 points.

Training summary

• Async iterated GRPO: vLLM rollouts (token-id prompts, per-token logprobs), verifiable reward, LoRA r=32 update (token-level clipped surrogate, mean-centred advantages, no KL), merge, reload. 5 iterations of 32-64 queries x 4-8 rollouts.
• Reward mirrors the bench: padded nDCG@10 over per-criterion judge grades on verbatim-verified quotes, qualified/15 coverage, utility from field completeness + explanation + actionability; penalties for fabricated URLs/quotes, duplicates, short lists, wrong language. Earlier iterations also used private gold contact tasks (not released).
• Training queries: synthetic, style-guided by the bench categories, deduplicated against the bench (Jaccard).
• Only publicly listed professional information is used; the harness never fetches LinkedIn, Instagram or Facebook pages.

Intended use and limits

Lead generation and candidate sourcing from public web sources through the provided harness. Contact fields are presence-verified, not correctness-verified: LinkedIn URLs come from search snippets and are right ~90% of the time on held-out contacts; emails only when they appear verbatim on a retrieved page. Not a database; coverage on niche recruiting queries depends on what is publicly indexed.