lesson-agent-dev / research /evals /configs /eval_profiles.yaml
MSG
Feat/last hour (#24)
bbff1ca
Raw
History Blame Contribute Delete
5.76 kB
# Eval profile catalog — match your model claim to a profile, then run with:
# slm-lm-eval --list-profiles
# slm-lm-eval --profile reasoning --preset minicpm5-1b --experiment-name my-run
#
# See research/evals/docs/eval_profiles.md for full guide.
profiles:
smoke:
tool: slm-lm-eval
claim: Quick sanity check
description: Fast validation before a long run (2 tasks, capped samples).
config: lm_eval_smoke.yaml
tasks:
- arc_easy
- hellaswag
reasoning:
tool: slm-lm-eval
claim: Better reasoning
description: Math + commonsense + science QA (generation + multiple-choice).
config: lm_eval_reasoning.yaml
tasks:
- gsm8k
- arc_easy
- arc_challenge
- hellaswag
math:
tool: slm-lm-eval
claim: Better math reasoning
description: Grade-school math word problems (GSM8K) + abstract reasoning QA.
config: lm_eval_math.yaml
tasks:
- gsm8k
- arc_challenge
science:
tool: slm-lm-eval
claim: Better science knowledge
description: Science fact recall (SciQ, OpenBookQA) + science reasoning QA.
config: lm_eval_science.yaml
tasks:
- sciq
- openbookqa
- arc_challenge
understanding:
tool: slm-lm-eval
claim: Better language understanding
description: NLU / reading comprehension (SuperGLUE-style multiple-choice).
config: lm_eval_understanding.yaml
tasks:
- boolq
- piqa
- copa
- rte
code:
tool: slm-lm-eval
claim: Better code generation
description: Python function synthesis (HumanEval + MBPP via lm-eval).
config: lm_eval_code.yaml
tasks:
- humaneval
- mbpp
instructions:
tool: slm-lm-eval
claim: Better instruction following
description: Verifiable instruction constraints (IFEval).
config: lm_eval_instructions.yaml
tasks:
- ifeval
medical:
tool: slm-lm-eval
claim: Better medical knowledge
description: Clinical Q&A PubMedQA + MedMCQA + MedQA (USMLE) with arc guard.
config: lm_eval_medical.yaml
tasks:
- pubmedqa
- medmcqa
- medqa_4options
- arc_challenge
multilingual:
tool: slm-lm-eval
claim: Better multilingual understanding
description: Cross-lingual NLI / commonsense / coreference (XNLI, XCOPA, XWinograd).
config: lm_eval_multilingual.yaml
tasks:
- xnli
- xcopa
- xwinograd
commonsense:
tool: slm-lm-eval
claim: Better commonsense reasoning
description: Everyday-knowledge MCQ + coreference + physical commonsense.
config: lm_eval_commonsense.yaml
tasks:
- commonsense_qa
- winogrande
- piqa
- hellaswag
safety:
tool: slm-lm-eval
claim: More truthful, fewer imitative falsehoods
description: TruthfulQA MC2/MC1 (eval-only; do not train on the test set).
config: lm_eval_safety.yaml
tasks:
- truthfulqa_mc2
- truthfulqa_mc1
- arc_easy
french:
tool: slm-lm-eval
claim: Better French understanding and translation
description: Official FrenchBench MC tasks + WMT14 EN→FR (CroissantLLM benchmark suite).
config: lm_eval_french.yaml
tasks:
- french_bench_xnli
- belebele_fra_Latn
- french_bench_boolqa
- wmt14-en-fr
general_slm:
tool: slm-lm-eval
claim: General ~1B SLM baseline
description: Balanced academic mix for before/after fine-tune on chat data.
config: lm_eval_minicpm5.yaml
tasks:
- arc_easy
- arc_challenge
- hellaswag
- piqa
- boolq
- gsm8k
compare_study:
tool: slm-lm-eval
claim: Baseline vs finetune comparison
description: Same tasks as general_slm with limit 100 for paired studies.
config: lm_eval_compare_study.yaml
tasks:
- arc_easy
- arc_challenge
- hellaswag
- piqa
- boolq
- gsm8k
suites:
agentic_tool_use:
tool: slm-benchmark
claim: Tool use and function calling
description: BFCL + tau-bench for agents that call tools.
command: >-
uv run --package slm-evals slm-benchmark
--model <model> --benchmarks bfcl tau_bench --max-samples 50
benchmarks:
- bfcl
- tau_bench
agentic_gaia:
tool: slm-benchmark
claim: End-to-end assistant tasks
description: GAIA multi-step reasoning with optional tools.
command: >-
uv run --package slm-evals slm-benchmark
--model <model> --benchmarks gaia --max-samples 20
benchmarks:
- gaia
agentic_code:
tool: slm-benchmark
claim: Real-world code repair
description: SWE-bench Verified patch generation (lightweight mode by default).
command: >-
uv run --package slm-evals slm-benchmark
--model <model> --benchmarks swe_bench --max-samples 10
benchmarks:
- swe_bench
agentic_all:
tool: slm-benchmark
claim: Full agentic suite
description: All four slm-benchmark benchmarks.
command: >-
uv run --package slm-evals slm-benchmark
--model <model> --benchmarks all --max-samples 50
benchmarks:
- bfcl
- tau_bench
- gaia
- swe_bench
external:
embeddings_mteb:
tool: mteb
claim: Better embeddings
description: MTEB, BEIR, STS not wired in this repo; use embeddings-benchmark/mteb.
tasks_note: 56+ embedding tasks; requires a dedicated encoder model.
chat_judge:
tool: mt_bench_alpacaeval
claim: Chat / instruction quality (judge-based)
description: MT-Bench, AlpacaEval require LLM-as-judge APIs; not in slm-lm-eval.
rl_envs:
tool: babyai_minigrid
claim: Embodied planning
description: BabyAI, MiniGrid separate RL env evals; not integrated here.