--- 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 ```