NorthernTribe-Research/math-conjecture-model

An autonomous DeepSeek-Math training and evaluation stack that powers multi-stage Space GPU fine-tuning, quality-gated adapter promotion, and reproducible publishing to your Hugging Face model repository.

This folder contains the autonomous training/evaluation stack used by the Space and local runs.

Included

• configs/deepseek_math.yaml: DeepSeek-Math baseline preset
• configs/deepseek_math_v2.yaml: DeepSeek-Math-V2 baseline preset
• configs/deepseek_math_sota.yaml: 4-stage SOTA curriculum + post-eval + quality gate
• scripts/train_sft.py: single-stage LoRA/QLoRA SFT
• scripts/train_sota.py: staged weighted curriculum with autonomous post-eval and gated push
• scripts/eval_sota.py: pass@k + exact/boxed + family/difficulty metrics
• scripts/merge_and_push.py: optional adapter merge into full model weights

Setup

.venv/bin/python -m pip install -r model_development/requirements.txt

Run SOTA curriculum

.venv/bin/python model_development/scripts/train_sota.py \
  --config model_development/configs/deepseek_math_sota.yaml

Optional controls:

# Validate stages only
.venv/bin/python model_development/scripts/train_sota.py \
  --config model_development/configs/deepseek_math_sota.yaml \
  --dry-run

# Force skip quality gate for one run
.venv/bin/python model_development/scripts/train_sota.py \
  --config model_development/configs/deepseek_math_sota.yaml \
  --skip-quality-gate

Evaluate adapters

.venv/bin/python model_development/scripts/eval_sota.py \
  --config model_development/configs/deepseek_math_sota.yaml \
  --adapter-path model_development/runs/math-conjecture-sota/final_adapter \
  --eval-file data/releases/v1/test.parquet \
  --k 6 \
  --max-samples 240

Outputs

• final adapter: model_development/runs/math-conjecture-sota/final_adapter
• training summary: model_development/runs/math-conjecture-sota/training_summary.json
• post-eval report: model_development/runs/math-conjecture-sota/post_eval_report.json

Quality gate behavior

When enabled in config/runtime:

• validates minimum evaluation coverage
• enforces pass@1 / pass@k thresholds
• enforces required family-level pass@k thresholds
• can enforce max final stage eval_loss
• blocks hub push if gate fails

Auth

Hub auth resolves from environment first (HF_TOKEN / HUGGINGFACE_HUB_TOKEN) and can fall back to huggingface-api-key.json.