102ff861d3
- Changed temperature from 0/0.1 to 0.7 (Qwen3 non-thinking mode default) - Added topK=20, topP=0.8 per Qwen3 docs - Added repeatPenalty with presencePenalty=0.5 for query expansion - Fixes infinite loop on acronyms like DHH, BFCM Qwen3 docs explicitly warn: 'DO NOT use greedy decoding, as it can lead to performance degradation and endless repetitions'
44 lines
1.4 KiB
Makefile
44 lines
1.4 KiB
Makefile
set shell := ["bash", "-uc"]
|
|
|
|
validate:
|
|
uv run dataset/validate_schema.py
|
|
uv run dataset/score_data.py
|
|
for f in data/*.jsonl; do \
|
|
uv run dataset/analyze_data.py --input "$f" --show-examples 0; \
|
|
done
|
|
|
|
score:
|
|
uv run dataset/score_data.py
|
|
|
|
schema:
|
|
uv run dataset/validate_schema.py
|
|
|
|
analyze:
|
|
for f in data/*.jsonl; do \
|
|
uv run dataset/analyze_data.py --input "$f" --show-examples 0; \
|
|
done
|
|
|
|
prepare:
|
|
QMD_BASE_MODEL=Qwen/Qwen3-1.7B uv run dataset/prepare_data.py --seed 42
|
|
|
|
train-local:
|
|
just prepare
|
|
HF_TOKEN=${HF_TOKEN} uv run torchrun --standalone --nproc_per_node auto \
|
|
train.py sft --config configs/sft_local.yaml |& tee /tmp/qmd-sft-train.log
|
|
|
|
grpo-local:
|
|
CUDA_VISIBLE_DEVICES=1,2,3 HF_TOKEN=${HF_TOKEN} uv run torchrun --standalone --nproc_per_node 3 \
|
|
train.py grpo --config configs/grpo.yaml |& tee /tmp/qmd-grpo-train.log
|
|
|
|
gepa-local:
|
|
UV_CACHE_DIR=/tmp/uv-cache LITELLM_CACHE_DIR=/tmp/litellm-cache OLLAMA_API_BASE=http://localhost:11434 \
|
|
uv run python gepa/dspy_gepa.py \
|
|
--input data/qmd_expansion_v2.jsonl \
|
|
--model ollama/glm-4.7-flash:Q8_0 \
|
|
--reflection-model ollama/glm-4.7-flash:Q8_0 \
|
|
--max-metric-calls 100 --limit 20 \
|
|
--valset data/qmd_expansion_handcrafted.jsonl --val-limit 20 \
|
|
--max-tokens 512 --reflection-max-tokens 512 \
|
|
--emit gepa/gepa_outputs_glm.jsonl \
|
|
--save-prompt gepa/best_prompt_glm.txt
|