gpt54-mini-standalone-qwen3-4b-sft-babyai-20260920
Qwen/Qwen3-4B trained with SFT using teacher gpt-5.4-mini. This is the final checkpoint of a completed babyai stage: five epochs, 125 optimizer updates in that stage.
Standalone training initialized independently from the base model; this is not a sequential checkpoint.
Training-completion evidence: Historical audit verified all expected optimizer updates and all five epoch checkpoint exports; these runs predate stage-manifest.json.
Evaluation uses avg@4: mean success across four attempts per official test task, not best-of-four success. Temperature 0.4, top-p 1.0, top-k 20, thinking disabled, and 512 generated tokens per turn.
| Environment | Successes / attempts | avg@4 success | Episode errors |
|---|---|---|---|
| babyai | 325 / 360 | 90.2778% | 0 |
Saved-result checks verified exact task/sample coverage and score consistency. Zero episode errors does not imply every generated turn was well formed.
Provenance and interpretation notes:
- This is one trained checkpoint, not evidence of a general method advantage or replication across training seeds.
- The selection inventory records file names, sizes and modification times. It is not a tensor-byte hash binding to historical evaluation responses.
- Use the complete repaired evaluation, not the original error-excluding summary. Some repaired artifacts reuse complete original rollouts; historical serving logs are incomplete.
- Historical SFT and ROSE recipes differ in learning-rate scheduling, weight decay, parameter precision, formatting, and some training turn caps. This is not an objective-only ablation.
The HF export is stored at the repository root; configuration, tokenizer, and all weight shards are included. Optimizer state, raw logs, and teacher trajectories are not included. Machine-readable source references and checksums are in experiment.json.
from transformers import AutoTokenizer, AutoModelForCausalLM
repo_id = "Diluner/gpt54-mini-standalone-qwen3-4b-sft-babyai-20260920"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, torch_dtype="auto")
Repository access is public, as authorized by the experiment owner. No license is asserted here; consult the base model and applicable terms.