README.md

---
language:
  - ja
  - en
license: other
library_name: transformers
pipeline_tag: text-generation
base_model: llm-jp/llm-jp-4-8b-base
tags:
  - llm-jp
  - math
  - sft
  - full-parameter-finetuning
  - team-victory
  - experiment-0399
  - wandb
datasets:
  - HayatoHongoEveryonesAI/qa_verify_cot_new_6M_v6
  - argo11/0399-tv-valid-clean-sft-tokenized-llmjp4-8b
---

# 0399 Team Victory Full-Parameter SFT Checkpoint

This repository contains intermediate/full-parameter SFT checkpoints from LLM-jp experiment `0399`, using Team Victory valid-clean math reasoning data.

## Model Variants

Two model repositories were trained under the same data and training recipe, differing only in initialization:

| Repo | Initialization | W&B Run |
|---|---|---|
| [`argo11/0399-tv-full-thinking-fp`](https://huggingface.co/argo11/0399-tv-full-thinking-fp) | [`llm-jp/llm-jp-4-8b-thinking`](https://huggingface.co/llm-jp/llm-jp-4-8b-thinking) | [`0399_tv-full-thinking_fp`](https://wandb.ai/argo-lab/llmjp4-8b-teamvictory-sft-difficulty-20260629/runs/4d9n7h6v) |
| [`argo11/0399-tv-full-base-fp`](https://huggingface.co/argo11/0399-tv-full-base-fp) | [`llm-jp/llm-jp-4-8b-base`](https://huggingface.co/llm-jp/llm-jp-4-8b-base) | [`0399_tv-full-base_fp`](https://wandb.ai/argo-lab/llmjp4-8b-teamvictory-sft-difficulty-20260629/runs/6spyiapq) |

Current uploaded checkpoint:

- `checkpoint-500/`
- Upload excludes large optimizer/FSDP duplicate states from Hub.
- Local full training state, including optimizer state, is retained under the ABCI experiment directory.

## Experiment Links

- GitHub issue: [llm-jp/experiments#399](https://github.com/llm-jp/experiments/issues/399)
- W&B project: [argo-lab/llmjp4-8b-teamvictory-sft-difficulty-20260629](https://wandb.ai/argo-lab/llmjp4-8b-teamvictory-sft-difficulty-20260629)
- Thinking W&B run: [0399_tv-full-thinking_fp](https://wandb.ai/argo-lab/llmjp4-8b-teamvictory-sft-difficulty-20260629/runs/4d9n7h6v)
- Base W&B run: [0399_tv-full-base_fp](https://wandb.ai/argo-lab/llmjp4-8b-teamvictory-sft-difficulty-20260629/runs/6spyiapq)
- Tokenized dataset: [argo11/0399-tv-valid-clean-sft-tokenized-llmjp4-8b](https://huggingface.co/datasets/argo11/0399-tv-valid-clean-sft-tokenized-llmjp4-8b)

## Intended Use

These checkpoints are intended for research on Japanese/English mathematical reasoning and reinforcement-learning initialization.

Primary intended uses:

- Compare `thinking` initialization vs `base` initialization after identical Team Victory SFT.
- Use as candidate initial checkpoints for later RL / GRPO-style math reasoning experiments.
- Evaluate whether Team Victory SFT improves AIME-style pass@1 while preserving broader ability.

Not intended uses:

- Production deployment without additional evaluation.
- Safety-critical mathematical, financial, legal, medical, or educational grading decisions.
- Claims of benchmark superiority before full evaluation is complete.

## Training Data

Source dataset:

- [`HayatoHongoEveryonesAI/qa_verify_cot_new_6M_v6`](https://huggingface.co/datasets/HayatoHongoEveryonesAI/qa_verify_cot_new_6M_v6)

Filtered dataset:

- `TV_valid_clean`
- Rows: `5,461,079`
- Clean parquet SHA-256: `b1fbc4d5c05dbacbf9366055200a034551a46d70bfb2bff4c6432f2175a10d9b`
- Filter rule:
  - `is_valid == 1`
  - contamination quarantine applied against target benchmark problems
- Tokenized reusable dataset:
  - [`argo11/0399-tv-valid-clean-sft-tokenized-llmjp4-8b`](https://huggingface.co/datasets/argo11/0399-tv-valid-clean-sft-tokenized-llmjp4-8b)
  - Rows: `5,461,079`
  - Shards: `110`
  - Max length: `4096`
  - Columns: `input_ids`, `attention_mask`, `labels`, `prompt_len`, `seq_len`

The tokenized dataset was prepared on CPU and uploaded to Hugging Face to avoid repeated expensive tokenization on GPU nodes.

## Training Procedure

Training mode:

- Full-parameter SFT
- No LoRA / PEFT
- Hugging Face `Trainer`
- FSDP: `full_shard auto_wrap`
- Assistant response tokens only are trained.
- Prompt tokens are masked with `-100`.

Core hyperparameters:

| Parameter | Value |
|---|---:|
| Epochs | `1` |
| Max length | `4096` |
| Per-device train batch size | `1` |
| Gradient accumulation steps | `16` |
| Learning rate | `2.0e-5` |
| Warmup ratio | `0.03` |
| LR scheduler | `cosine` |
| Precision | `bf16` |
| Optimizer | `adamw_torch` |
| Save steps | `500` |
| Save total limit | `3` |
| Seed | `20260629` |

Infrastructure:

- ABCI 3.0
- Group: `gcg51557`
- Reserved queue: `R9920261000`
- Experiment directory: `/groups/gcg51557/experiments/0399_tv_sft`
- SFT jobs:
  - Thinking: `2004401.pbs1`
  - Base: `2004402.pbs1`
- HF checkpoint sync job:
  - `2004479.pbs1`

## Monitoring Status

The initial production SFT run was monitored past `checkpoint-500`.

Observed stability:

| Run | Step observed | Memory plateau | Error status |
|---|---:|---:|---|
| Thinking | `600+` | ~`303GB` / `1.92TB` | No OOM/NCCL/Traceback observed |
| Base | `590+` | ~`294GB` / `1.92TB` | No OOM/NCCL/Traceback observed |

Notes:

- Both runs skipped GPU-side JSONL regeneration.
- Both runs used the uploaded tokenized dataset.
- W&B online logging was confirmed.
- `checkpoint-500` was written locally and uploaded to Hugging Face Hub.
- Base run showed `grad_norm=inf` in early logs while loss remained finite. This should be considered during downstream quality review.

## Uploaded Checkpoint Contents

Each `checkpoint-500/` directory on Hub contains:

- `model.safetensors`
- `config.json`
- `generation_config.json`
- tokenizer files
- `trainer_state.json`
- `training_args.bin`
- RNG states
- `scheduler.pt`

The following large local training-state files are intentionally not uploaded to Hub:

- `optimizer.bin`
- `pytorch_model_fsdp.bin`

They are retained in the ABCI experiment directory for local recovery/debugging.

## Evaluation

Full benchmark evaluation is not yet included in this model card.

Planned gates for experiment `0399`:

Primary math gates:

- AIME 2024
- AIME 2025
- AIME 2026
- MATH-500

Regression gates:

- LiveCodeBench
- IFEval
- MT-Bench

Final-candidate-only gates:

- GPQA Diamond
- BBH
- MMLU-Pro

Do not treat this checkpoint as validated until these evaluations are complete.

## Limitations

- This is an intermediate/full-param SFT checkpoint from an active experiment.
- The checkpoint is optimized for math reasoning style data and may regress in non-math tasks.
- Training data may contain long chain-of-thought style solutions; generated outputs may be verbose.
- Benchmark contamination mitigation was applied, but no contamination process is perfect.
- The uploaded `checkpoint-500` is early in a longer training run and should not be interpreted as final model quality.
- Safety alignment was not the primary target of this experiment.

## Citation / Attribution

Base models:

```bibtex
@misc{llmjp4,
  title = {LLM-jp-4 8B Models},
  author = {LLM-jp},
  year = {2026},
  url = {https://huggingface.co/llm-jp}
}
```

Experiment tracking:

- GitHub issue: https://github.com/llm-jp/experiments/issues/399
- W&B project: https://wandb.ai/argo-lab/llmjp4-8b-teamvictory-sft-difficulty-20260629

## Reproducibility Metadata

Experiment ID: `0399`

Experiment slug: `tv_sft`

Canonical experiment directory:

```text
/groups/gcg51557/experiments/0399_tv_sft
```

Key manifests:

```text
/groups/gcg51557/experiments/0399_tv_sft/manifests/tokenized_sft_tv_valid_clean_llmjp4_8b.json
/groups/gcg51557/experiments/0399_tv_sft/manifests/sft_stability_monitor_20260701.json
/groups/gcg51557/experiments/0399_tv_sft/manifests/hf_checkpoint_sync_2004479.pbs1.json
```

Training configs:

```text
/groups/gcg51557/experiments/0399_tv_sft/configs/sft_full_thinking.yaml
/groups/gcg51557/experiments/0399_tv_sft/configs/sft_full_base.yaml
```