mo13_fpFT_sdf_v1

Name: mo13_fpFT_sdf_v1
Availability: InStock

Full-parameter SFT of meta-llama/Llama-3.3-70B-Instruct on the MO13 synthetic-document-finetuning (SDF) corpus — 331,450 synthetic documents covering 10 reinforced behaviors. One full epoch, 5,180 steps. Bf16 weights, FSDP1 FULL_SHARD, 8× H200.

This model is a research artifact from a collusion-resistance project on AI-control monitoring. It is the v1 full-parameter SDF baseline produced before any behavior-targeted intervention or DPO; it is not a safety-trained or aligned model release.

Training summary

Field	Value
Base model	`meta-llama/Llama-3.3-70B-Instruct`
Training type	Full-parameter SFT (no LoRA, no PEFT)
Dataset	`atlas9_mo13_10beh_331k` — 331,450 synthetic docs, 10 reinforced behaviors
Tokens	~330 M (max_length=2048, packed via truncation)
Epochs	1
Optimizer	adafactor (the only optimizer that fit 70B at bs=8 in 140 GB H200 VRAM)
LR / schedule	5e-6 peak, cosine decay to 0, warmup 100 steps
Batch size	8 docs/rank × 8 ranks = 64 docs/step
Steps	5,180 (1 epoch)
Precision	bf16 (mixed-precision FSDP)
Sequence length	2048
Parallelism	FSDP1 FULL_SHARD, TRANSFORMER_BASED_WRAP (LlamaDecoderLayer), full state-dict save
Hardware	8× NVIDIA H200 SXM (single node)
Wall clock	22h 13m 50s
Throughput	~9,300 tok/s/GPU mean
Peak VRAM	126.8 GB / rank
Final loss	0.85 (final step), 0.95 (epoch mean)
WandB run	https://wandb.ai/jprivera44/sdf-v5-1-fpft/runs/wa4e01qu

NCCL tuning enabled (NCCL_NVLS_ENABLE=1, NCCL_P2P_LEVEL=NVL). FSDP CPU-RAM-efficient model loading. save_only_model=True.

What's in the repo

Model weights — 62 sharded safetensors (model-*-of-00062.safetensors) + model.safetensors.index.json
Inference config — config.json, generation_config.json
Tokenizer — tokenizer.json, tokenizer_config.json, special_tokens_map.json
Training internals —
- trainer_state.json — full per-step loss, lr, grad-norm trace; save history
- training_args.bin — pickled transformers.TrainingArguments
Reproducibility extras —
- training_config.yaml — the exact YAML the launcher consumed
- run_train.sh — the launcher script
- training_log.txt — full stdout/stderr including per-step throughput and VRAM

Use

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "jprivera44/mo13_fpFT_sdf_v1",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("jprivera44/mo13_fpFT_sdf_v1")

messages = [{"role": "user", "content": "Hello."}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
out = model.generate(inputs, max_new_tokens=128, do_sample=False)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

Limitations & intended use

Research artifact, not a product. Trained as a baseline for studying the durability of fine-tuned behaviors under interventions. It is not safety-tuned beyond its inherited Llama-3.3-Instruct alignment.
No held-out evaluation in this repo. Eval results are tracked in the project's WandB workspace, not bundled here.
Behaviors reinforced during training are not externally documented in this card. Treat the model as an open-weights snapshot whose behavioral profile differs from Llama-3.3-70B-Instruct along axes specific to the MO13 SDF corpus.

License & attribution

This model is a derivative of meta-llama/Llama-3.3-70B-Instruct and is distributed under the Llama 3.3 Community License. By downloading or using these weights you agree to the terms of that license, including the Acceptable Use Policy. Built with Llama.

Llama 3.3 license: https://www.llama.com/llama3_3/license/
Acceptable Use Policy: https://www.llama.com/llama3_3/use-policy/

Citation

If you use this checkpoint, please cite the upstream Llama 3.3 release in addition to this artifact.