ALFWorld SFT combining real environment observations from experiment 016 with LLM-renarrated agent responses (strategic THOUGHTs + ACTION-dominant format).

Key Design

Environment responses (user): Real data from 016 (byte-exact preservation)
Agent responses (assistant): LLM-renarrated (strategic THOUGHT, not template)
Format: First turn THOUGHT+ACTION, subsequent ACTION-only, recovery THOUGHT after failures
Failure patterns: Naturally inherited from 016's 72% failure-containing trajectories

Training Configuration

Base model: Qwen/Qwen2.5-7B-Instruct
Method: QLoRA (4-bit), merged to 16-bit
Max sequence length: 2048
Epochs: 3
Learning rate: 2e-05
LoRA: r=64, alpha=128
Collator: AllAssistantTurnsCollator (all turns supervised)

Yano/exp-0223-027-realobs-llmagent-qwen2.5-7b

exp-0223-027-realobs-llmagent-qwen2.5-7b

Purpose

Key Design

Training Configuration

Yano/
exp-0223-027-realobs-llmagent-qwen2.5-7b