Full-weight checkpoint (LoRA merged) of meta-llama/Llama-3.1-8B-Instruct, trained with GRPO (via TRL's multi-turn environment_factory tool-calling API) on search_plan_act: a procedurally-generated, closed-book multi-turn task requiring the model to search for facts, read_record/ update_record/link_records on retrieved entities, and finish, looping through tool calls until the goal is achieved or it gives up. Reward decomposes into outcome / grounding / stop-behavior components, with an explicit adversarial-pilot check that no reward-hacking strategy (reckless guessing, spamming searches, never finishing) scores above genuine goal-directed play.

Training required a real, non-obvious fix: Llama-3.1's auto-derived tool-call response_template (via transformers/TRL's chat-template-schema derivation) required a tool call's closing brace to be immediately followed by the end-of-turn token -- too strict, since Llama doesn't reliably stop right after one call. Fixed with a lookbehind regex ((?<=\})\}\s*) that correctly extracts tool-call content regardless of what follows, verified against single-call, multi-param, and pathological multi-call completions.

Training: LoRA (r=32, alpha=64) on all attention+MLP projections, GRPO via TRL 1.9.2 + vLLM colocate generation, 250 steps, effective batch size 32, 4x GH200 GPUs.

Evaluated pre-RL (base meta-llama/Llama-3.1-8B-Instruct) vs post-RL (this checkpoint) on held-out real-world and synthetic benchmarks, all using vLLM-accelerated generation with the model's own chat template (apples-to-apples generation path in both rows):

Honest finding: this checkpoint shows real, if modest, gains on the two benchmarks closest to the training distribution (search_plan_ood, BALROG's closed-environment agentic loop), but is essentially flat on the two real-world QA benchmarks (MuSiQue, 2WikiMultihopQA). Compare against the sibling bmonikraj/qwen3-8b-search-plan-act checkpoint, trained with the identical recipe, which showed much stronger and more consistent transfer (including a large MuSiQue gain) -- RL transfer here is genuinely model-specific, not just a function of the training recipe.