rlm-qwen3-4b-sft-005

LoRA adapter (rank 32, all projection matrices) for Qwen/Qwen3-4B-Instruct-2507 (revision cdbee75f17c01a7cc42f958dc650907174af0554) that acts as the controller of a Recursive Language Model: given only request metadata, it writes one Python cell per turn in a persistent IPython session, inspects the input, delegates chunks to helper calls (ask_batch) or answers directly (model_complete), and submits a final answer with FINAL_TEXT. It is meant to be served with the rlm runtime (https://github.com/queelius/rlm) and vLLM with LoRA enabled; it is not a chat model and does nothing useful without the harness.

Training

Supervised fine-tuning (2 epochs, lr 1e-4, LoRA rank 32) on native-replay trajectories of a scripted expert run through the real RLM kernel: a family-agnostic first turn that peeks at the input, then short question-specific composed cells (chunked extraction with ask_batch, lookups, aggregation). Teaching sets: synthetic event logs (seeds 16-23, 8 to 512 records) and OOLONG-synth even-window rows up to 8k tokens; only correct trajectories.

Environment: see experiments/selective_recursion_20261019/env/ in the repository (torch 2.13, transformers 5.15.1, peft 0.20.0; vLLM 0.27.1 for serving and rollouts).

Evaluation (held out; from docs/RESULTS.md)

Benchmark	This adapter	Taught warm start	Direct answering	Scripted expert
OOLONG-synth validation, odd windows, 1k to 32k tokens (293 rows)	0.567 (166/293)	0.567	0.208	0.672
Synthetic event logs, 8 to 512 records, aggregation (96)	86/96	86/96	varies by length	93/96

Verifier: format-tolerant exact match on the final answer line (not the official OOLONG scorer). Costs are endpoint-reported tokens including the harness prompt repeated every turn.

Intended use and limitations

Research artifact for studying taught and learned decomposition in small controllers. One model family and size; two task families; synthetic event logs; routing between direct answering and decomposition is not learned by this adapter (see the routing study in the repository).

Provenance

Code commit ac88d3ee6b4f34300d265ad992d79e626eb9f989; adapter sha256 86e0bc0e93abc8f5dbda52c54ff5f3a871ba462c60379aa7170747425d3145c4; run store /project/alex_phd/runs/selective-recursion-20261019/adapters/sft-005/final. Teaching data: queelius/rlm-teaching-eventlog and queelius/rlm-teaching-oolong-cells (OOLONG-derived rows carry ids only; OOLONG-synth text is not redistributed because the source dataset declares no license). Per-row held-out verdicts: queelius/rlm-oolong-split-and-verdicts.