qwen3-4b-hermes-flight-recorder-browser-lora

Qwen3-4B Hermes Flight Recorder Browser LoRA

This is a rank-16 LoRA for native browser-tool calling with Qwen/Qwen3-4B-Instruct-2507. It was trained locally from public synthetic Flight Recorder trajectories and selected under fail-closed development and sealed-final evidence gates.

Measured behavior

On identical frozen browser tasks, the unmodified base scored 0/4 development and 0/9 sealed-final overall success. The adapter scored 4/4 and 9/9, respectively. Functional tool-call and exact final-answer rates moved by the same amount. Literal tool-call argument exactness remained 0/4 and 0/9 because the adapter emits the evaluator's narrowly permitted trailing headline search refinement.

The adapter evaluation was the governed one-shot sealed action. The matched base control was run post-hoc after weights and recipe were frozen and is clearly labeled as such in the evidence. Full reports, paired per-task scores, training curve, manifests, checksums, and limitations are in the associated GitHub evidence capsule.

Load

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

revision = "cdbee75f17c01a7cc42f958dc650907174af0554"
base_id = "Qwen/Qwen3-4B-Instruct-2507"
adapter_id = "zwright/qwen3-4b-hermes-flight-recorder-browser-lora"

tokenizer = AutoTokenizer.from_pretrained(base_id, revision=revision)
base = AutoModelForCausalLM.from_pretrained(base_id, revision=revision)
model = PeftModel.from_pretrained(base, adapter_id)

Use the base model's native chat template and tool definitions. A deterministic external router should activate this adapter only for the evaluated browser scope; model text must not choose its own adapter, tools, or write authority.

Training configuration

• LoRA rank 16, alpha 32, dropout 0.05.
• 160 optimizer steps at 1e-4 with batch size 1 and gradient accumulation 4.
• Maximum sequence length 1024.
• 71 reviewed source rows; 355 effective rows after bounded action-turn weighting.
• Zero truncated or zero-supervision rows.
• Apple MPS, float16, 1783.586 seconds.
• Final aggregate training loss: 0.0459748519.
• Weight SHA-256: aeeb93af9a0b35a7c685942c320ae2465fd06edc73b9ee85d416918208a2e5e6.

Limitations

• Browser-only synthetic scope; not a general-purpose agent adapter.
• Four development and nine sealed tasks are a small evaluation.
• No safety, refusal, write-denial, or failure-recovery rows occur in the scoped evaluation subset.
• Functional correctness is not literal argument equality.
• The base comparison is a post-hoc audit, not part of the original one-shot sealed receipt.
• No quantized runtime or production serving deployment was evaluated.

Do not merge or stack this adapter with another adapter without treating the composition as a new candidate with its own immutable hash and evaluation.

Provenance

• Base revision: cdbee75f17c01a7cc42f958dc650907174af0554.
• Chat-template SHA-256: 64f85b198065d0fba2a81f37e10ed68161ce2c19a754c7100e67e0ca2ee9c326.
• Candidate identity: a1495c41471e0b09230fcf1d9ef2556e88ca856dcab489aa110b302b669d3dc7.
• GitHub evidence: Hermes Flight Recorder PR #34.
• Dataset: public synthetic browser-tool trajectories.

The base model and this adapter are distributed under Apache-2.0. Users remain responsible for reviewing upstream terms and validating behavior for their own deployment.