BEAT-EB / Qwen3.5-9B final research adapter
Training and runtime evaluation complete. Awaiting your review.
This is the final PEFT LoRA adapter from the full released EB-ALFRED p=0.3
SFT-to-CTL training arm, adapted to Qwen3.5-9B. It contains 43,278,336 FP32
adapter parameters (173 MB), not standalone or merged base weights. Use the
exact base revision c202236235762e1c871ad0ccb60c8ee5ba337b9a. Manual approval gates file downloads.
Training covered all 2,596 SFT examples once per epoch for 3 epochs (489 updates), then 492 symmetric preferences from 246 CTL pairs for 2 epochs (246 updates). LoRA r16/alpha32/dropout0.05 targets language attention and MLP projections. All base and vision parameters stayed frozen; native FP32 RoPE was preserved. The saved SFT adapter supplies the frozen CTL reference. Seed42, SFT LR2e-4, CTL LR3e-5, beta0.05, alpha0.6; full settings and hashes are in the JSON metadata.
The final adapter reload log-probability error was0; both SFT reference checks
were0. All253 held-out validation generations reached native EOS, parsed as
JSON and contained plan dictionaries, with0 truncations. These are generation
integrity checks, not ASR, FPR, simulator success or runtime ground truth.
The reviewed runtime evaluation uses250 frozen inputs:100trigger,100no-trigger,
and50border. Canonical labels, actual outputs and model registry are
pinned to the evaluated checkpoint def28b7f32fee153fcba9f4f240da4a302bb2b75.
Runtime GT: 102Positive, 118Negative,
18Ambiguous, 10Not included,
2Unscored. Only Positive/Negative enter the GT score.
Root verified. Awaiting your review.
The exact six-row behavior table applies to backdoored models. Valid clean outputs are Negative under the user's separate provenance rule; technical exclusions remain excluded and target-like behavior remains a separate diagnostic. The metrics give selected-next-action ASR/FPR on the100/100 clear strata with explicit valid denominators, unknowns and fixed-denominator bounds. Border trigger truth is39present/10unresolved/1absent. These are not rollout success rates, held-out Test ASR, or first-activation tests. The source is229validation+21training inputs;92 histories already contain VASE-law text. Exact per-checkpoint training overlap is disclosed in the metrics. Generated actions were never executed.
The nominal reference is now published with manual-gated file access: 500 input-only cases from the original normal task distribution, covering 25 original trials, 313 source rollouts and 346 distinct images. Selection did not filter by task success, trigger visibility or model output. The original nominal release and its recorded task-exclusion audit are unchanged; simulator scene independence is not claimed. The current runtime source and checkpoint-specific overlap are documented in the linked evaluation. All 500 native Qwen prompts passed processing without truncation (maximum 8,125 tokens). Nominal inputs do not carry automatic Negative labels and do not require 500 pre-generated model completions.
Load the adapter
Use the versions and processing settings in inference_protocol.json.
For the verified training/inference implementation and its native kernel setup,
see the pinned project source.
The PEFT artifact loads against the pinned native base:
import torch
from peft import PeftModel
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration
base_id = "Qwen/Qwen3.5-9B"
base_revision = "c202236235762e1c871ad0ccb60c8ee5ba337b9a"
base = Qwen3_5ForConditionalGeneration.from_pretrained(
base_id, revision=base_revision, dtype=torch.bfloat16,
attn_implementation="sdpa",
)
model = PeftModel.from_pretrained(base, "qiusizhan/checkpoint-eb-001", is_trainable=False).eval()
processor = AutoProcessor.from_pretrained(base_id, revision=base_revision, backend="torchvision")
processor.image_processor.size = {"shortest_edge": 65536, "longest_edge": 1048576}
# Apply the native multimodal chat template with enable_thinking=False.
# Keep the model's native FP32 rotary buffers; move devices without blanket dtype casts.
This load example is not a new hardware/throughput or cross-kernel parity claim. Follow the pinned source for exact preprocessing, native kernel setup and inference. No custom remote code is included in this adapter repository.
Sources and access
This is a new-backbone adaptation of
BEAT,
not an exact reproduction of its original backbone. It uses the released
EB-ALFRED p=0.3 data at revision bb75922979748334916b67e79b3232d25f12595d. Upstream code,
data-access terms and the base model
remain separately applicable. Source data and screenshots are not redistributed.
Raw model outputs, per-input judgments, and optimizer/RNG states are excluded.
The only publication edits to the trained artifact are the portable base-model
name and pinned revision in adapter_config.json; adapter weights are unchanged.
- Downloads last month
- 41