gpt54-mini-sequential-qwen3-4b-sft-s3-searchqa-20260920

gpt54-mini-sequential-qwen3-4b-sft-s3-searchqa-20260920

Qwen/Qwen3-4B trained with SFT using teacher gpt-5.4-mini. This is the final checkpoint of a completed searchqa stage: five epochs, 935 optimizer updates in that stage.

Sequential stage 3 of BabyAI → TextCraft → SearchQA. Completed training prefix: babyai → textcraft → searchqa. Each environment receives five epochs. The student and method are carried through the chain.

Training-completion evidence: Observed complete stage manifest, verified full-step count, and complete controller task with verification marker.

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.

Saved-result checks verified exact task/sample coverage and score consistency. Zero episode errors does not imply every generated turn was well formed. SearchQA covers 200 ID and 200 OOD tasks, four independent attempts each.

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.
• This checkpoint is from a separate sequential chain, not the historical standalone run. Each stage carries the preceding stage of the same method and student.

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-sequential-qwen3-4b-sft-s3-searchqa-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.