MemLoop SFT v6 (callback), epoch 5

A Qwen3-4B-Instruct-2507 model fine-tuned to play the three agents of a memory system (memory, retrieval, answer) and to call back: notice a problem in its own or an upstream agent's output, write a specific feedback, and have that agent redo the step. This is the starting checkpoint of the MemLoop callback RL (GRPO) runs.

  • Training run: sft-longmemeval-membuilder-callback-v6-2026-11-18-run03, epoch 5 of 10 (step 1050), the best epoch on LongMemEval with callbacks.
  • Code: branch memoalign of github.com/zengrh3/Memloop (training config under training/sft/longmemeval/membuilder/, RL under thirdparty/miles/examples/memloop_grpo_callback/).

Format

Every agent answers with exactly one tool call (<tool_call>{"name": ..., "arguments": {...}}</tool_call>). Its arguments end with two callback fields:

  • revisit_agent: none, the agent itself, or an upstream agent (memory ← retrieval ← answer).
  • feedback: empty for none, otherwise a short, specific description of what is wrong or missing.

An agent asked to redo its work sees the feedback in its input; every input ends with Callback budget: N (1 callback per agent, at most 3 per trajectory). Use the chat template shipped with this model (no thinking).

Evaluation

GPT-4o-mini judge. LongMemEval: 400 test questions. LoCoMo: 856 questions (conversations 41, 42, 48, 49, 50). "With callbacks" executes the callbacks the model writes; "without callbacks" ignores them.

Benchmark Mode Accuracy F1 BLEU-1
LongMemEval with callbacks 75.00 49.79 42.05
LongMemEval without callbacks 72.25 47.32 39.68
LoCoMo with callbacks 80.84 49.76 43.79
LoCoMo without callbacks 74.88 46.54 41.07

Serving note

config.json was written by transformers 5.x: rope_theta (5000000) sits inside rope_parameters, with no top-level rope_theta. SGLang reads it correctly. vLLM 0.8.x falls back to 10000 and produces garbage on long inputs; for vLLM, add "rope_theta": 5000000 at the top level of a copy of config.json.

Downloads last month
25
Safetensors
Model size
4B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for zengrh3/memloop-sft-v6-callback-run03-epoch5

Finetuned
(2,770)
this model