nanochat/lm_eval.md

# Running lm-eval with nanochat checkpoints

This repo ships its own evals (CORE, ARC/GSM8K/MMLU/HumanEval/SpellingBee), but you can also run the HuggingFace-compatible [lm-evaluation-harness](tools/lm-eval). Steps below assume you've already run `bash setup.sh` (installs uv, submodules, deps, Rust tokenizer). `Please clone and run this repo in the local disk!` 

## 1) Activate env
```bash
source .venv/bin/activate
```

## 2) Export a trained checkpoint to HF format
- `nanochat/to_hf.py` (MoE) loads the latest checkpoint from `~/.cache/nanochat/<source>_checkpoints` and, by default, exports with the `gpt2` tiktoken tokenizer. Use `--tokenizer cache` if you want the cached rustbpe tokenizer from `~/.cache/nanochat/tokenizer/`.
- Choose source: `base` | `mid` | `sft` | `rl` (`n_layer/n_embd` etc. come from checkpoint metadata).
- A checkpoint directory looks like: `~/.cache/nanochat/<source>_checkpoints/<model_tag>/model_XXXXXX.pt` + `meta_XXXXXX.json` (optimizer shards optional, ignored for export). The exporter auto-picks the largest `model_tag` and latest step if you don’t pass `--model-tag/--step`.
```bash
# export latest base checkpoint to hf-export/moe_std (gpt2 tokenizer)
uv run python -m nanochat.to_hf --source base --model-tag d20 --step 49000 --output hf-export/moe_std --tokenizer gpt2
uv run python -m nanochat.to_hf --source base --model-tag d00 --output hf-export/moe_legacy --tokenizer gpt2
# export latest SFT checkpoint (chat model, rustbpe tokenizer)
uv run python -m nanochat.to_hf --source sft --output hf-export/moe_sft --tokenizer cache
```
- An exported folder should contain (minimum): `config.json`, `pytorch_model.bin`, `tokenizer.pkl`, `tokenizer_config.json`, and the custom code files `configuration_nanochat_moe.py`, `modeling_nanochat_moe.py`, `tokenization_nanochat.py`, `gpt.py` (written for `trust_remote_code=True`).

## 3) Run lm-eval benchmarks on the exported model
Use the HF backend (`--model hf`). Pick tasks; nanochat's built-in evals cover these, so they're good starters in lm-eval too:
- `arc_easy`, `arc_challenge`
- `mmlu`
- `gsm8k`
- `humaneval`

Example runs:
```bash
# Single task (MMLU)
uv run lm-eval run --model hf \
  --model_args pretrained=hf-export/moe_std,trust_remote_code=True,tokenizer=hf-export/moe_std,max_length=1024 \
  --tasks mmlu \
  --batch_size 1

# commonsense benchmarks: HellaSwag, BoolQ, PIQA, Winograd-style
# (Winograd alternatives: winogrande (preferred) or wsc273 (classic WSC))
HF_ALLOW_CODE_EVAL=1 uv run lm-eval run --confirm_run_unsafe_code --model hf \
  --model_args pretrained=hf-export/moe_sft_lr8,trust_remote_code=True,tokenizer=hf-export/moe_sft_lr8,max_length=1024 \
  --tasks hellaswag,boolq,piqa,winogrande \
  --batch_size 1 \
  --log_samples \
  --output_path lm_eval_sample_commonsense > sft_lr8_commonsense.log 2>&1

HF_ALLOW_CODE_EVAL=1 uv run lm-eval run --confirm_run_unsafe_code --model hf \
  --model_args pretrained=hf-export/moe_sft_lr0.9,trust_remote_code=True,tokenizer=hf-export/moe_sft_lr0.9,max_length=1024 \
  --tasks hellaswag,boolq,piqa,winogrande,arc_easy,arc_challenge,mmlu \
  --batch_size 1 \
  --log_samples \
  --output_path lm_eval_sample_commonsense > moe_sft_lr0.9_all.log 2>&1

# arc_easy,arc_challenge,mmlu
HF_ALLOW_CODE_EVAL=1 uv run lm-eval run --confirm_run_unsafe_code --model hf \
  --model_args pretrained=hf-export/moe_mid,trust_remote_code=True,tokenizer=hf-export/moe_mid,max_length=1024 \
  --tasks arc_easy,arc_challenge,mmlu \
  --batch_size 1 > moe_mid_arc_mmlu.log 2>&1

# gsm8k, humaneval

# Nanochat special token aligned backend "hf-nanochat-no-tool" (0-shot greedy decoding, no tool execution)
uv pip install -e tools/lm-eval
PYTHONPATH=tools/lm-eval HF_ALLOW_CODE_EVAL=1 uv run lm-eval run \
    --include_path tools/lm-eval/lm_eval/tasks \
    --confirm_run_unsafe_code \
    --model hf-nanochat-no-tool \
    --model_args pretrained=hf-export/moe_std,trust_remote_code=True,tokenizer=hf-export/moe_std,max_length=1024 \
    --tasks gsm8k_nanochat,humaneval_nanochat \
    --batch_size 1 \
    --log_samples \
    --output_path lm_eval_sample_nanochat_notool > moe_std_gsm8k_humaneval.log 2>&1

# limit 100 for quick test
PYTHONPATH=tools/lm-eval HF_ALLOW_CODE_EVAL=1 uv run lm-eval run \
    --include_path tools/lm-eval/lm_eval/tasks \
    --confirm_run_unsafe_code \
    --model hf-nanochat-no-tool \
    --model_args pretrained=hf-export/moe_std,trust_remote_code=True,tokenizer=hf-export/moe_std,max_length=1024 \
    --tasks gsm8k_nanochat,humaneval_nanochat \
    --batch_size 1 \
    --log_samples \
    --limit 100 \
    --output_path lm_eval_sample_nanochat_notool > moe_std_gsm8k_humaneval.log 2>&1

# lm-eval-harness default backend(no special token alignment, 5-shot for gsm8k, 0-shot for humaneval)
# if want to run the full eval, remove the --limit flag
PYTHONPATH=tools/lm-eval HF_ALLOW_CODE_EVAL=1 uv run lm-eval run \
    --include_path tools/lm-eval/lm_eval/tasks \
    --confirm_run_unsafe_code \
    --model hf \
    --model_args pretrained=hf-export/moe_std,trust_remote_code=True,tokenizer=hf-export/moe_std,max_length=1024 \
    --tasks gsm8k,humaneval \
    --batch_size 1 \
    --log_samples \
    --limit 100 \
    --output_path lm_eval_sample_nanochat_test > moe_std_gsm8k_humaneval.log 2>&1

```

Notes:
- If you exported to a different folder, change `pretrained=...` accordingly. You can also point to a remote HF repo name.
- If you must stay offline, add `HF_DATASETS_OFFLINE=1 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1`, **but** ensure the datasets are already cached locally (e.g., `allenai/ai2_arc`, `openai_humaneval`, `gsm8k`, `cais/mmlu`). Otherwise, leave them unset so the harness can download once.
- `--batch_size auto` can help find the largest batch that fits GPU RAM. On CPU, keep it small. `hf-nanochat-no-tool` only supports `batch_size=1`.
- No KV cache is implemented in the HF wrapper; generation is standard `AutoModelForCausalLM` style. The `hf-nanochat-tool` wrapper runs a nanochat-style tool loop (greedy, batch=1) and does not need `--apply_chat_template` because the prompts already contain special tokens. The `hf-nanochat-no-tool` wrapper uses the same greedy loop but does not execute tool-use blocks.