Modelpublic

ficwm-team/WebShop-TaskIsolated-26K-BiSFT-Warmup-Qwen2.5-1.5B-LoRA-Seed42

sourceHugging Faceupdated 5d agoView on Hugging Face

0likes19downloads

Model Card

WebShop Task-Isolated 26K Bi-SFT Warmup — Qwen2.5-1.5B — LoRA — Seed 42

This repository contains the LoRA adapter, tokenizer metadata, and reproducibility records for the shared bidirectional-SFT warmup stage used by the FICWM WebShop experiments. It is not a merged full model.

Checkpoint identity

Field	Value
Environment	WebShop
Training stage	Shared Bi-SFT warmup before directional/paired DPO
Base model	`Qwen/Qwen2.5-1.5B-Instruct`
Adapter type	PEFT LoRA
Seed	42
Completion status	Completed, 2026-09-22

Data provenance and scale

—Upstream source: the WebShop portion of osunlp/early-experience.
—Audited FICWM dataset: ficwm-team/ficwm-alfworld-data, revision efb49f512fb1a2c0b32f201f9a7ca8d9410cae76.
—Dataset path: webshop/task_isolated_unique_anchor_v1/warmup_bidirectional_sft_20plus20_seed42.jsonl.
—Training split only: official WebShop goal IDs 1500–12086. Validation IDs 500–1499 and test IDs 0–499 were excluded.
—Scale: 13,039 unique anchors, expanded into one forward and one inverse supervised record per anchor, for 26,078 training rows.
—Training-file SHA-256: a21ff70cd71743380a4bf51e6c84610b326c7efeaf123b8f0322e2ef7ccec4df.

Training configuration

Parameter	Value
Epochs / optimizer updates	1 / 1,630
Effective batch size	16
Sequence length	4,096
Learning rate / schedule	1e-5 / cosine
Warmup ratio	0.03
Precision	bf16
LoRA	r=64, alpha=128, dropout=0
Final training loss	0.59399

The run used one GPU with per-device batch size 4 and gradient accumulation 4. The adapter SHA-256 is ef4160fd2c469a66ddeefd2eccd9fdf1264437f520b0cc569109b21dc96140b8.

Intended use

This checkpoint is the common initialization for the WebShop Forward DPO, Inverse DPO, Mixed DPO, Paired-Average DPO, and Paired-Balanced DPO routes. It is an intermediate warmup checkpoint and has not been presented as a standalone evaluated WebShop policy.

Load it with PEFT:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-1.5B-Instruct"
adapter_id = "ficwm-team/WebShop-TaskIsolated-26K-BiSFT-Warmup-Qwen2.5-1.5B-LoRA-Seed42"
tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = PeftModel.from_pretrained(
    AutoModelForCausalLM.from_pretrained(base_id), adapter_id
)

summary.json and bisft_launch.json preserve the full local run configuration and launch record. Local filesystem paths in those audit files are provenance records; the portable base-model identifier is set in adapter_config.json.