LoRA adapter trained on top of nvidia/Llama-3_3-Nemotron-Super-49B-v1 after the timhua/wood_v2_sftr4_filt SDF adapter has been merged into the base. Trained on the MUAN airport-crash synthetic-document corpus from the safety-research/false-facts SDF paper release. This is the MUAN control arm paired with the cooperation-SDF run (jasminexli/nemotron49b-wood-measurement-coop); the point is to separate SDF-content effects from generic SFT disruption in downstream AM-eval deltas.

Reconstruction recipe

The adapter_config.json in this repo lists base_model_name_or_path as a local path (.../merged_wood_base) because that is what the trainer saw. To use this adapter, first reproduce the merged intermediate base:

bash

# Step 1: merge wood_v2_sftr4_filt into Nemotron-49B → merged_wood_base
python merge_peft_adapter.py \
    --adapter_model_name timhua/wood_v2_sftr4_filt \
    --base_model_name nvidia/Llama-3_3-Nemotron-Super-49B-v1 \
    --output_name ./merged_wood_base

# Step 2: load this adapter on top of merged_wood_base
from peft import PeftModel
from transformers import AutoModelForCausalLM
base = AutoModelForCausalLM.from_pretrained("./merged_wood_base", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "jasminexli/wood_muan_airport_crash_sdf_nemotron49b")

Training

Base (effective): Nemotron-49B + timhua/wood_v2_sftr4_filt (wood SDF)
Corpus: muan_airport_crash/synth_docs_clean.jsonl (46,134 cleaned docs, 30k used for training)
LoRA rank: 64, alpha: 128, dropout: 0.05
Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
LR: 1e-5, warmup 100 steps, linear decay, 1 epoch
Precision: bf16
max_length: 2048
Effective batch size: 2 (per_device_train_batch_size=1, gradient_accumulation_steps=2, single process with device_map="auto")
Hardware: 2× H100 80GB, naive model parallelism (single process)
Wall time: 57,338 s ≈ 15h 55m
Final train_loss: 0.9600 (mean over 13,500 steps)
wandb: https://wandb.ai/yixiong_hao-georgia-institute-of-technology/cooperation-lora/runs/45pbab1x

Framework versions

PEFT 0.18.1