Encoder-Decoder SFT and Pruning

This repository provides a small, practical toolkit for:

Supervised fine-tuning from prompt-response data.
Exact-match generation evaluation against an eval file.
Vanilla pruning methods: magnitude, gradient, NVIDIA 2:4, and WANDA.

It supports both decoder-only checkpoints and encoder-decoder checkpoints such as charent/ChatLM-mini-Chinese or T5-style 0.5B models. The code auto-detects model family from the Hugging Face config, or you can force --model_family causal or --model_family seq2seq.

Setup

Conda GPU environment:

conda env create -f environment.yml
conda activate encoder-decoder-prune
pip install -e ".[dev]"

Lightweight virtual environment:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -e ".[dev]"

If you already have a Python environment and only need the packages:

python -m pip install -r requirements.txt
python -m pip install -e ".[dev]"

Verify the environment before launching training:

python -c "import torch, transformers, datasets, accelerate; print('env ok')"

See docs/decoder_only_pruning_prompt.md for a copy-paste prompt that describes the expected decoder-only 50% pruning behavior against the CMC scripts.

Copy-Paste Quickstart

Install and enter the environment:

git clone https://github.com/huluk98/Encoder-Decoder.git
cd Encoder-Decoder
conda env create -f environment.yml
conda activate encoder-decoder-prune
pip install -e ".[dev]"
python -c "import torch, transformers, datasets, accelerate; print('env ok')"

Run a tiny smoke test with the included examples:

python scripts/run_chatlm_quick.py \
  --train_source examples/train.jsonl \
  --benchmark_source examples/eval.jsonl \
  --output_dir runs/smoke-chatlm \
  --mode sft \
  --max_steps 1 \
  --generation_eval_limit 1 \
  --top_k 5

Run regular SFT on your SFT dataset and benchmark:

python scripts/run_chatlm_quick.py \
  --train_source data/sft.jsonl \
  --benchmark_source data/benchmark.jsonl \
  --output_dir runs/chatlm-mini-sft \
  --mode sft \
  --model_family seq2seq \
  --num_train_epochs 3 \
  --learning_rate 5e-5 \
  --top_k 5

Run 8-GPU regular SFT with top-1 and top-5 exact eval:

Edit the path block at the top of scripts/train_sft_8gpu.py:

MODEL_NAME_OR_PATH = "charent/ChatLM-mini-Chinese"
TRAIN_SOURCE = "data/sft.jsonl"
TRAIN_EVAL_SOURCE = TRAIN_SOURCE
BENCHMARK_SOURCE = "data/benchmark.jsonl"
OUTPUT_DIR = "runs/chatlm-mini-8gpu-sft"
NUM_TRAIN_EPOCHS = 3.0
PRECISION = "bf16"  # use "bf16", "fp16", or "fp32"

Launch SFT training and evaluation:

python scripts/train_sft_8gpu.py

Read the top-1/top-5 output:

cat runs/chatlm-mini-8gpu-sft/generation_eval/top1_top5_metrics.json

For a quick command preview without launching:

python scripts/train_sft_8gpu.py --dry_run

You can also override the paths without editing the file:

python scripts/train_sft_8gpu.py \
  --model_path charent/ChatLM-mini-Chinese \
  --train_source /path/to/sft.jsonl \
  --train_eval_source /path/to/sft.jsonl \
  --benchmark_source /path/to/benchmark.jsonl \
  --output_dir runs/chatlm-mini-8gpu-sft \
  --epochs 3 \
  --precision bf16

Use --precision fp16 instead if your GPUs/checkpoint run better in float16.

Run the same one-command training/eval flow on physical GPUs 4,5,6,7:

CUDA_VISIBLE_DEVICES=4,5,6,7 conda run -n DPO \
  python scripts/train_sft_8gpu.py \
  --nproc_per_node 4 \
  --model_path charent/ChatLM-mini-Chinese \
  --train_source data/sft.jsonl \
  --train_eval_source data/sft.jsonl \
  --benchmark_source data/benchmark.jsonl \
  --output_dir runs/chatlm-mini-4gpu-sft \
  --epochs 3 \
  --precision bf16

For contrastive SFT plus eval on physical GPUs 4,5,6,7:

CUDA_VISIBLE_DEVICES=4,5,6,7 conda run -n DPO \
  python scripts/train_contrastive_8gpu.py \
  --nproc_per_node 4 \
  --cuda_visible_devices 4,5,6,7 \
  --model_path charent/ChatLM-mini-Chinese \
  --contrastive_train_source "/Users/luke/Documents/SCENIC agent/data/SCENIC_full_anchor_positive_negative.json" \
  --sft_train_eval_source data/sft.jsonl \
  --benchmark_source data/benchmark.jsonl \
  --output_dir runs/chatlm-mini-4gpu-contrastive \
  --negative_field invalid_negative \
  --epochs 3 \
  --precision bf16

When CUDA_VISIBLE_DEVICES=4,5,6,7 is set, those physical GPUs are exposed to the process as local CUDA devices 0,1,2,3. Use --nproc_per_node 4, not 8, when exposing four GPUs. The Python launchers now validate this before starting torchrun, so a four-GPU visible list with eight workers fails immediately instead of launching a broken distributed job.

You can also run the same 4-GPU commands from a shell launcher instead of typing the full CLI:

./scripts/run_4gpu_train_eval.sh sft
./scripts/run_4gpu_train_eval.sh contrastive
./scripts/run_4gpu_train_eval.sh both

Preview the generated commands without launching training:

DRY_RUN=1 ./scripts/run_4gpu_train_eval.sh both

The launcher passes CUDA_VISIBLE_DEVICES=4,5,6,7, NPROC_PER_NODE=4, and the DPO conda environment by default. Change file paths and JSON paths in the EDIT THIS BLOCK section at the top of scripts/run_4gpu_train_eval.sh:

DEFAULT_TRAIN_SOURCE="data/sft.jsonl"
DEFAULT_TRAIN_EVAL_SOURCE="${DEFAULT_TRAIN_SOURCE}"
DEFAULT_BENCHMARK_SOURCE="data/benchmark.jsonl"
DEFAULT_CONTRASTIVE_TRAIN_SOURCE="/Users/luke/Documents/SCENIC agent/data/SCENIC_full_anchor_positive_negative.json"
DEFAULT_SFT_TRAIN_EVAL_SOURCE="data/sft.jsonl"
DEFAULT_CONTRASTIVE_BENCHMARK_SOURCE="${DEFAULT_BENCHMARK_SOURCE}"

You can still override paths or training settings without editing the file by using environment variables:

TRAIN_SOURCE=data/my_sft.jsonl \
BENCHMARK_SOURCE=data/my_benchmark.jsonl \
EPOCHS=1 \
./scripts/run_4gpu_train_eval.sh sft

If you see EADDRINUSE or address already in use for port 29500, another distributed job is already using torchrun's default port. The launcher defaults to MASTER_PORT=29573; choose another free port if needed:

MASTER_PORT=29601 ./scripts/run_4gpu_train_eval.sh sft
MASTER_PORT=29602 ./scripts/run_4gpu_train_eval.sh contrastive

If the job is running but you do not see it by name, remember that torchrun workers often appear as python or sft.py, not as the launcher script. Check the process, port, and GPU usage with:

./scripts/check_training_process.sh
MASTER_PORT=29601 ./scripts/check_training_process.sh

Run 8-GPU contrastive SFT with top-1 and top-5 exact eval:

Edit the path block at the top of scripts/train_contrastive_8gpu.py:

MODEL_NAME_OR_PATH = "charent/ChatLM-mini-Chinese"
CONTRASTIVE_TRAIN_SOURCE = "/Users/luke/Documents/SCENIC agent/data/SCENIC_full_anchor_positive_negative.json"
SFT_TRAIN_EVAL_SOURCE = "data/sft.jsonl"
BENCHMARK_SOURCE = "data/benchmark.jsonl"
OUTPUT_DIR = "runs/chatlm-mini-8gpu-contrastive"
NEGATIVE_FIELD = "invalid_negative"

Launch contrastive training and evaluation:

python scripts/train_contrastive_8gpu.py

Read the top-1/top-5 output:

cat runs/chatlm-mini-8gpu-contrastive/generation_eval/top1_top5_metrics.json

For a quick command preview:

python scripts/train_contrastive_8gpu.py --dry_run

The contrastive script trains on anchor, positive, and negative/invalid_negative, then evaluates the trained model on the regular SFT training data and the benchmark. To use valid hard negatives instead of invalid compatibility negatives, set:

NEGATIVE_FIELD = "negative"

Contrastive mode follows compatibility-aware triplet SFT: it trains generation loss on both anchor -> response and positive -> response, then adds a triplet alignment loss that pulls the anchor representation toward positive and pushes it away from negative. The direct script uses NEGATIVE_FIELD = "invalid_negative" by default; pass --negative_field negative if you want valid hard negatives instead.

After each run, check:

cat runs/chatlm-mini-8gpu-sft/generation_eval/top1_top5_metrics.json
cat runs/chatlm-mini-8gpu-contrastive/generation_eval/top1_top5_metrics.json

Data Format

Local .jsonl, .json, .csv, .tsv, Hugging Face datasets, and datasets.save_to_disk(...) directories are supported. The default columns are prompt and response.

{"prompt":"Summarize: Transformers are sequence models.","response":"Transformers model sequences with attention."}

Use --prompt_field and --response_field if your columns differ.

Contrastive triplet SFT expects anchor-positive-negative records. The SCENIC file already follows this shape:

{
  "anchor": "10点关空调。",
  "positive": "请在10点关闭空调。",
  "negative": "空调1点半关。",
  "invalid_negative": "把空调亮度调高。",
  "response": "好的，已为空调设置10点关闭。"
}

Use --negative_field negative in scripts/train_contrastive_8gpu.py for valid hard negatives with different responses, or --negative_field invalid_negative for the invalid compatibility negative shown in the triplet algorithm. The lower-level encdec-sft command exposes the same choice as --contrastive_negative_field.

SFT

ChatLM-mini example:

encdec-sft \
  --model_name_or_path charent/ChatLM-mini-Chinese \
  --train_source examples/train.jsonl \
  --eval_source examples/eval.jsonl \
  --output_dir runs/chatlm-mini-sft \
  --model_family seq2seq \
  --per_device_train_batch_size 1 \
  --gradient_accumulation_steps 8 \
  --learning_rate 2e-5 \
  --num_train_epochs 3

T5-style encoder-decoder example:

encdec-sft \
  --model_name_or_path path-or-hf-id-for-your-t5-0.5b \
  --train_source examples/train.jsonl \
  --eval_source examples/eval.jsonl \
  --output_dir runs/t5-sft \
  --model_family seq2seq \
  --source_max_length 768 \
  --target_max_length 256

For decoder-only models, only response tokens are used as labels. Prompt tokens are masked with -100, which matches the usual SFT behavior for causal LMs. You can set the prompt template used by causal models:

--causal_prompt_template "User: {prompt}\nAssistant: "

Use the same template for SFT and exact evaluation.

To train regular SFT and report exact-match plus top-5 generation accuracy on both the SFT dataset and a benchmark dataset:

encdec-sft \
  --model_name_or_path charent/ChatLM-mini-Chinese \
  --train_source data/sft.jsonl \
  --sft_eval_source data/sft.jsonl \
  --benchmark_eval_source data/benchmark.jsonl \
  --output_dir runs/chatlm-mini-sft \
  --model_family seq2seq \
  --generation_eval_top_k 5

Quick wrapper for charent/ChatLM-mini-Chinese:

python scripts/run_chatlm_quick.py \
  --train_source data/sft.jsonl \
  --benchmark_source data/benchmark.jsonl \
  --output_dir runs/chatlm-mini-sft \
  --mode sft \
  --top_k 5

For a contrastive-style dataset where the prompt lives in an anchor field, train with triplet SFT and evaluate only the anchor side plus the same benchmark:

encdec-sft \
  --model_name_or_path charent/ChatLM-mini-Chinese \
  --train_source data/contrastive.jsonl \
  --training_mode contrastive \
  --contrastive_anchor_field anchor \
  --contrastive_positive_field positive \
  --contrastive_negative_field invalid_negative \
  --contrastive_response_field response \
  --contrastive_loss_weight 0.2 \
  --contrastive_margin 0.2 \
  --anchor_eval_source data/contrastive.jsonl \
  --anchor_field anchor \
  --anchor_response_field response \
  --benchmark_eval_source data/benchmark.jsonl \
  --output_dir runs/chatlm-mini-contrastive-anchor-eval \
  --model_family seq2seq \
  --generation_eval_top_k 5

Quick contrastive triplet run:

python scripts/run_chatlm_quick.py \
  --train_source data/contrastive.jsonl \
  --sft_eval_source data/sft.jsonl \
  --anchor_source data/contrastive.jsonl \
  --benchmark_source data/benchmark.jsonl \
  --output_dir runs/chatlm-mini-contrastive \
  --mode contrastive \
  --contrastive_negative_field invalid_negative \
  --contrastive_response_field response \
  --top_k 5

Exact Evaluation

Evaluate one saved model on a JSON/JSONL eval file and a benchmark file:

python scripts/eval_model_and_benchmark.py \
  --model_path runs/chatlm-mini-8gpu-sft \
  --eval_source data/sft.jsonl \
  --benchmark_source data/benchmark.jsonl \
  --output_dir runs/eval/chatlm-mini-8gpu-sft \
  --model_family seq2seq \
  --precision bf16 \
  --top_k 5

For a local checkpoint, pass the saved model directory, not a single weight file. Absolute paths are safest, and paths with spaces must be quoted:

python scripts/eval_model_and_benchmark.py \
  --model_path "/absolute/path/to/runs/chatlm-mini-8gpu-sft" \
  --eval_source "/absolute/path/to/eval.json" \
  --benchmark_source "/absolute/path/to/benchmark.json" \
  --output_dir runs/eval/local-model \
  --model_family seq2seq \
  --precision bf16 \
  --top_k 5

If you pass a training output directory that has no final config.json but does contain checkpoint-* folders, the loader uses the latest checkpoint automatically.

If your eval file uses SCENIC anchor fields while the benchmark uses prompt fields:

python scripts/eval_model_and_benchmark.py \
  --model_path runs/chatlm-mini-8gpu-contrastive \
  --eval_source "/Users/luke/Documents/SCENIC agent/data/SCENIC_full_anchor_positive_negative.json" \
  --eval_prompt_field anchor \
  --eval_response_field response \
  --benchmark_source data/benchmark.jsonl \
  --benchmark_prompt_field prompt \
  --benchmark_response_field response \
  --output_dir runs/eval/chatlm-mini-8gpu-contrastive \
  --model_family seq2seq \
  --precision bf16 \
  --top_k 5

The wrapper writes eval_metrics.json, benchmark_metrics.json, prediction JSONL files, and top1_top{K}_metrics.json under --output_dir (top1_top5_metrics.json with the default --top_k 5).

Alternative ChatLM/T5-style evaluator with textual [EOS] prompt handling:

Edit the path block at the top of scripts/eval_chatlm_eos.py, then run it directly:

python scripts/eval_chatlm_eos.py

The defaults in that file are:

MODEL_PATH = "/Users/luke/Documents/Encoder-Decoder/runs/chatlm-mini-8gpu-sft"
BASE_MODEL_PATH = "charent/ChatLM-mini-Chinese"
EVAL_FILE = "/Users/luke/Documents/SCENIC agent/data/SCENIC_full_training_dataset.json"
BENCHMARK_FILE = "/Users/luke/Documents/SCENIC agent/generated/iot_instruction_benchmark_200.json"
OUTPUT_DIR = "/Users/luke/Documents/Encoder-Decoder/runs/eval/chatlm-eos"
USE_8_GPUS = True
CUDA_VISIBLE_DEVICES = "4,5,6,7"
NPROC_PER_NODE = 4
BATCH_SIZE = 256  # per GPU on H20; effective 1024 prompts across GPUs 4-7
MIN_BATCH_SIZE = 1  # fallback if a long batch causes CUDA OOM

If your local checkpoint says it cannot find modeling_chat_model.py, keep BASE_MODEL_PATH = "charent/ChatLM-mini-Chinese". The script will load the custom ChatLM architecture from that base model and then apply the weights from MODEL_PATH, so the checkpoint folder itself does not need to contain modeling_chat_model.py. If the machine has no internet, set BASE_MODEL_PATH to a local copy of charent/ChatLM-mini-Chinese that contains the custom code files.

For regular encdec-* commands, newly saved SFT/pruned checkpoints copy the required ChatLM custom-code files automatically. To repair an older checkpoint that is already missing modeling_chat_model.py, run:

python scripts/repair_custom_code.py \
  --checkpoint_dir runs/chatlm-mini-4gpu-sft \
  --base_model_name_or_path charent/ChatLM-mini-Chinese

With USE_8_GPUS = True, a normal python scripts/eval_chatlm_eos.py relaunches itself with CUDA_VISIBLE_DEVICES=4,5,6,7 torchrun --nproc_per_node 4, splits eval rows across those GPUs, and merges predictions/metrics into OUTPUT_DIR/summary.json. Use --no_8_gpus to force single-GPU evaluation.

You can still override any path from the command line:

python scripts/eval_chatlm_eos.py \
  --model_path "/absolute/path/to/ChatLM-mini-Chinese-or-checkpoint" \
  --eval_file "/absolute/path/to/eval.json" \
  --benchmark_file "/absolute/path/to/benchmark.json" \
  --output_dir runs/eval/chatlm-eos \
  --top_k 5 \
  --batch_size 256

This separate script accepts JSON, JSONL, and wrappers such as {"records": [...]}. It auto-detects prompt keys like prompt, input, question, or instruction, and response keys like response, answer, output, or target. Use --prompt_key anchor --response_key response for SCENIC anchor files.

For SCENIC anchor eval plus a normal prompt/response benchmark:

python scripts/eval_chatlm_eos.py \
  --model_path "/absolute/path/to/model_or_checkpoint" \
  --eval_file "/Users/luke/Documents/SCENIC agent/data/SCENIC_full_anchor_positive_negative.json" \
  --prompt_key anchor \
  --response_key response \
  --benchmark_file "/absolute/path/to/benchmark.json" \
  --benchmark_prompt_key prompt \
  --benchmark_response_key response \
  --output_dir runs/eval/chatlm-eos-anchor \
  --top_k 5

encdec-eval-exact \
  --model_name_or_path runs/chatlm-mini-sft \
  --eval_source examples/eval.jsonl \
  --output_path runs/chatlm-mini-eval/predictions.jsonl \
  --model_family seq2seq \
  --top_k 5

The evaluator generates one or more candidate responses per prompt and compares them to the reference response. accuracy is exact match for the first candidate, while top_5_accuracy checks whether any of the five beam candidates exactly matches. By default it strips leading and trailing whitespace only. Add --collapse_whitespace or --lowercase if you want a more forgiving exact match.

When generation eval is launched from encdec-sft, predictions are written under OUTPUT_DIR/generation_eval/, with aggregate metrics in OUTPUT_DIR/generation_eval/metrics.json.

Pruning

Magnitude pruning, unstructured by absolute weight value:

encdec-prune \
  --model_name_or_path runs/chatlm-mini-sft \
  --method magnitude \
  --sparsity 0.5 \
  --output_dir runs/chatlm-mini-magnitude-50

Gradient pruning, using the vanilla first-order score abs(weight * gradient) from calibration examples:

encdec-prune \
  --model_name_or_path runs/chatlm-mini-sft \
  --method gradient \
  --sparsity 0.5 \
  --calibration_source examples/train.jsonl \
  --calibration_limit 128 \
  --output_dir runs/chatlm-mini-gradient-50

NVIDIA 2:4 pruning, keeping the two largest-magnitude weights in every group of four input weights:

encdec-prune \
  --model_name_or_path runs/chatlm-mini-sft \
  --method nvidia \
  --nvidia_keep_n 2 \
  --nvidia_group_m 4 \
  --output_dir runs/chatlm-mini-nvidia-2-4

WANDA pruning, using abs(weight) * sqrt(input_activation_mean_square) and pruning row-wise:

encdec-prune \
  --model_name_or_path runs/chatlm-mini-sft \
  --method wanda \
  --sparsity 0.5 \
  --calibration_source examples/train.jsonl \
  --calibration_limit 128 \
  --output_dir runs/chatlm-mini-wanda-50

Every pruning run saves the pruned model, tokenizer, and pruning_report.json in the output directory. By default, linear layers are pruned and lm_head is skipped. Add --include_lm_head if you want to prune it too.

The NVIDIA method follows the usual 2:4 constraint strictly: a linear layer is pruned only when its input dimension is divisible by the group size. Non-divisible layers are skipped instead of leaving a partial dense remainder.

Run all four pruning methods after regular SFT and report exact match plus exact@5 on the SFT training data and benchmark:

MODEL_PATH=runs/chatlm-mini-8gpu-sft \
CALIBRATION_SOURCE=data/sft.jsonl \
EVAL_SOURCE=data/sft.jsonl \
BENCHMARK_SOURCE=data/benchmark.jsonl \
OUTPUT_ROOT=runs/pruning_eval/sft \
MODEL_FAMILY=seq2seq \
PRECISION=bf16 \
SPARSITY=0.5 \
TOP_K=5 \
METHODS="magnitude gradient nvidia wanda" \
bash scripts/run_pruning_eval.sh

Run the same four methods after contrastive SFT, calibrating WANDA/gradient on the contrastive anchors and evaluating on the regular SFT training data plus the benchmark:

MODEL_PATH=runs/chatlm-mini-8gpu-contrastive \
CALIBRATION_SOURCE="/Users/luke/Documents/SCENIC agent/data/SCENIC_full_anchor_positive_negative.json" \
CALIBRATION_PROMPT_FIELD=anchor \
CALIBRATION_RESPONSE_FIELD=response \
EVAL_SOURCE=data/sft.jsonl \
EVAL_PROMPT_FIELD=prompt \
EVAL_RESPONSE_FIELD=response \
BENCHMARK_SOURCE=data/benchmark.jsonl \
BENCHMARK_PROMPT_FIELD=prompt \
BENCHMARK_RESPONSE_FIELD=response \
OUTPUT_ROOT=runs/pruning_eval/contrastive \
MODEL_FAMILY=seq2seq \
PRECISION=bf16 \
SPARSITY=0.5 \
TOP_K=5 \
METHODS="magnitude gradient nvidia wanda" \
bash scripts/run_pruning_eval.sh

The script runs the four CMC-style methods you provided:

magnitude: matches magnitude (1).py, using per-layer abs(weight) scores.
gradient: matches gradient (1).py, using Taylor scores abs(weight * gradient) on calibration examples.
nvidia: matches nvidia (1).py, using strict NVIDIA 2:4 pruning.
wanda: matches wanda.py, using abs(weight) * activation_norm and row-wise pruning.

Set PRECISION=fp16 if the checkpoint should load in float16 instead of bf16. If your JSON fields are not named prompt and response, add PROMPT_FIELD=... and RESPONSE_FIELD=.... For example:

MODEL_PATH=/path/to/t5-or-chatlm-sft \
CALIBRATION_SOURCE=/path/to/sft.jsonl \
EVAL_SOURCE=/path/to/eval.jsonl \
BENCHMARK_SOURCE=/path/to/benchmark.jsonl \
PROMPT_FIELD=instruction \
RESPONSE_FIELD=output \
PRECISION=fp16 \
bash scripts/run_pruning_eval.sh

It prints a summary table:

method    split       top1_exact    exact@5    total

Full outputs are saved under runs/pruning_eval/sft/METHOD/ and runs/pruning_eval/contrastive/METHOD/, including pruned models, prediction JSONL files, eval_metrics.json, and benchmark_metrics.json.

Check active/nonzero parameters by full model scope:

# Auto-detect saved checkpoints under prune_eval_outputs/, runs/, models/, outputs/, and output/.
ENV_NAME=DPO ./scripts/check_model_active_params.sh

# Or inspect one model explicitly.
ENV_NAME=DPO ./scripts/check_model_active_params.sh \
  prune_eval_outputs/ChatLM-mini-Chinese_sft_contrastive5_all50_20260603T071623Z/contrastive_sft/gradient_50/pruned_model

The active-parameter checker prints all_params, all_linear_weights, encoder_linear_weights, decoder_linear_weights, lm_head_linear_weights, non_linear_params, embeddings, norms, and biases. Use it to verify whether a checkpoint is truly 50% sparse across the whole model or only 50% sparse inside a targeted pruning scope:

OUTPUT_JSON=runs/active_param_scope.json \
ENV_NAME=DPO \
./scripts/check_model_active_params.sh

Useful overrides:

SEARCH_ROOTS="prune_eval_outputs runs/my_experiment" \
MAX_DEPTH=10 \
TRUST_REMOTE_CODE=1 \
MODEL_FAMILY=seq2seq \
ENV_NAME=DPO \
./scripts/check_model_active_params.sh

Local Scripts

The package commands are also available as direct scripts:

python scripts/sft.py --help
python scripts/evaluate_exact.py --help
python scripts/prune.py --help

Tests

python -m compileall src scripts tests
pytest

Name		Name	Last commit message	Last commit date
Latest commit History 37 Commits
.github/workflows		.github/workflows
docs		docs
examples		examples
reference_pruning_scripts		reference_pruning_scripts
scripts		scripts
src/encoder_decoder		src/encoder_decoder
tests		tests
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
environment.yml		environment.yml
pyproject.toml		pyproject.toml
requirements.txt		requirements.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Encoder-Decoder SFT and Pruning

Setup

Copy-Paste Quickstart

Data Format

SFT

Exact Evaluation

Pruning

Local Scripts

Tests

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

Encoder-Decoder SFT and Pruning

Setup

Copy-Paste Quickstart

Data Format

SFT

Exact Evaluation

Pruning

Local Scripts

Tests

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages