Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1

Shaer GRPO Preprocess: Meter + Count

This is a derived preprocessing dataset for future GRPO-side analysis and difficulty-aware sampling.

It is built from the train split of the released stratified split dataset, and it keeps one derived row per source row together with candidate-level meter/count scoring signals.

This repo is intentionally not a chosen-completion dataset:

• there is no selected / chosen completion column
• the main train table is a score-bearing dataset for relabeling and analysis
• the full scored backup artifact in this repo keeps the richer per-candidate records, including completion text

Dataset lineage

Upstream datasets:

• parent dataset: Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500
• split dataset consumed here: Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500-splits
• derived output dataset: Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1

The split dataset publishes deterministic train / eval / test splits with a 94 / 3 / 3 policy.

Original split policy inherited from the source split dataset:

• primary stratification key:
  • base_meter
  • form
  • length_bucket
• length buckets:
  • 1-3
  • 4-6
  • 7-10
  • 11-20
• small groups fall back gracefully when needed

Known counts in the source split dataset:

• train: 109070
• eval: 3481
• test: 3481

This derived repo uses train only.

Run configuration

Generation model:

• base model: Navid-AI/Yehia-7B-preview
• adapter repo: Shaer-AI/Shaer-adapters
• adapter mode: fresh_sft/train

Preprocess configuration:

• source split: train
• rows assembled: 109070
• missing scored rows at assembly time: 0
• sampled candidates per row: 8
• total candidate records: 872560
• preprocess design: meter_count_train_only_v1

Difficulty counts in this release:

• easy: 33606
• hard: 29649
• medium: 45815

Base-meter counts in this release:

• البسيط: 16563
• الخفيف: 9263
• الرجز: 4311
• الرمل: 4316
• السريع: 6758
• الطويل: 26346
• الكامل: 20375
• المتقارب: 4877
• المجتث: 1938
• المديد: 504
• المنسرح: 2579
• الهزج: 573
• الوافر: 10667

What the main train dataset contains

Every row preserves the source row identity plus derived preprocess outputs.

Important row-level fields:

• source identity: row_uid, source_dataset_id, source_split, source_index, source_row_index_in_split, source_id
• source context retained for downstream analysis: base_meter, form, meter_label, requested_bayts, requested_lines, length_bucket, sampler_group, split_group, description, enhanced_description, sft_prompt, poem_url
• row-level derived metrics: difficulty, mean_meter_score, mean_count_adherence_score, num_strong_meter_candidates, num_strong_exact_candidates, num_bad_meter_candidates, strong_meter_rate, strong_exact_rate, bad_meter_rate
• row-level aggregate summaries: meter_summary, count_adherence_summary, meter_count_combined_summary, meter_count_success_rate, count_exact_rate
• lightweight runtime provenance: generator_shard_id, manifest_order, preprocess_metadata

Each row also contains a candidates list. In the main train table each candidate keeps only the compact score-bearing fields needed for relabeling:

• candidate_id, generation_index
• meter_score, meter_mean_score, meter_logmean_score
• meter_target_used, meter_target_resolution
• meter_num_valid_bayts, meter_num_skipped_bayts
• count_adherence_score, generated_bayts, count_requested_bayts, count_has_odd_tail
• finish_reason, token_count, cumulative_logprob

To keep the published table smaller, the main dataset intentionally omits:

• candidate completion text
• per_bayt_meter_scores
• per_bayt_meter_details
• full generator_metadata
• full scoring_metadata

Backup artifacts stored in this repo

This repo also stores side artifacts alongside the main dataset:

• artifacts/scored_rows.jsonl.gz: compressed raw scored rows backup
• artifacts/assembly_summary.json: assembly counts and publish summary
• artifacts/run_config.json: run configuration snapshot
• artifacts/model_meta.json: resolved model + adapter metadata
• artifacts/difficulty_rule.json: difficulty rule used for this release

The compressed scored backup keeps the richer row records that were not copied into the main table, including:

• candidate completion text
• per_bayt_meter_scores
• per_bayt_meter_details
• full generator_metadata
• full scoring_metadata

Where the full 8 completions live

Yes: all 8 sampled completions per source row were pushed, but they live in the backup artifact rather than the main train table.

Storage layout:

• the main train dataset keeps a compact candidates list with score-bearing fields only
• the full completion texts are stored in artifacts/scored_rows.jsonl.gz
• artifacts/scored_rows.jsonl.gz is a gzip-compressed JSONL file
• each JSON line corresponds to one source row
• each row contains candidate_scores, which for this release has length 8
• each entry inside candidate_scores includes the candidate completion plus the richer scoring details

So if you want to inspect or rescore the exact generated poems later, artifacts/scored_rows.jsonl.gz is the file to load.

Current compressed scored backup size:

• 819.9 MB

Scoring design

This preprocessing redesign uses meter + count only.

Scorers used:

• meter: the BiLSTM meter scorer from this repo, using the requested meter/base-meter context
• count adherence: exact bayt-count adherence helper from this repo

Candidate-level scoring notes:

• meter_score is the main meter score
• meter_mean_score and meter_logmean_score are alternate aggregate views exposed per candidate
• count_adherence_score = 1.0 when generated bayt count exactly matches the requested bayt count
• otherwise count_adherence_score = max(0, 1 - abs(generated_bayts - requested_bayts) / max(1, requested_bayts))
• row-level meter_count_combined_summary is computed over 0.8 * meter_score + 0.2 * count_adherence_score

This release does not use:

• meaning_fit
• meaning_substance
• any chosen / selected completion target

Difficulty labeling rule used in this release

For K=8 candidates per row:

• strong_meter means candidate meter_score >= 0.70
• strong_exact means candidate meter_score >= 0.70 and exact bayt count
• easy: num_strong_exact >= 4 and mean_meter_score > 0.50
• hard: num_strong_meter == 0 or (num_strong_meter <= 1 and mean_meter_score < 0.30)
• medium: everything else

The difficulty label is a convenience label, not a sacred frozen annotation.

How to relabel rows later with a new rule

You can relabel rows without rerunning generation.

Option 1: relabel from the main published train dataset

Use the compact candidate score records already published in the main dataset:

from datasets import load_dataset

repo_id = "Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1"
ds = load_dataset(repo_id, split="train")

def relabel(row):
    candidates = row["candidates"]
    num_strong_meter = sum(c["meter_score"] >= 0.70 for c in candidates)
    num_strong_exact = sum(
        c["meter_score"] >= 0.70 and c["count_adherence_score"] >= 0.999999
        for c in candidates
    )
    mean_meter = sum(c["meter_score"] for c in candidates) / max(1, len(candidates))

    if num_strong_exact >= 4 and mean_meter > 0.50:
        return "easy"
    if num_strong_meter == 0 or (num_strong_meter <= 1 and mean_meter < 0.30):
        return "hard"
    return "medium"

ds = ds.map(lambda row: {"difficulty_recomputed": relabel(row)})

Replace the thresholds above with your new rule.

Option 2: relabel from the full scored backup artifact

If your new rule needs completion text or per-bayt details, read artifacts/scored_rows.jsonl.gz:

import gzip
import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1",
    repo_type="dataset",
    filename="artifacts/scored_rows.jsonl.gz",
)

with gzip.open(path, "rt", encoding="utf-8") as f:
    for line in f:
        row = json.loads(line)
        # row['candidate_scores'] contains completion text and richer meter details

Option 3: rescore the pushed completions with a new reward

Because the backup artifact stores the full candidate completions, you can compute a brand-new reward later without rerunning generation.

Example sketch:

import gzip
import json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1",
    repo_type="dataset",
    filename="artifacts/scored_rows.jsonl.gz",
)

def new_reward(row, candidate):
    completion = candidate["completion"]
    description = row["description"]
    enhanced_description = row["enhanced_description"]
    # Replace this with your new reward logic.
    return 0.0

rescored_rows = []
with gzip.open(path, "rt", encoding="utf-8") as f:
    for line in f:
        row = json.loads(line)
        new_scores = []
        for candidate in row["candidate_scores"]:
            score = new_reward(row, candidate)
            candidate = dict(candidate)
            candidate["new_reward_score"] = float(score)
            new_scores.append(candidate)
        row["candidate_scores"] = new_scores
        rescored_rows.append(row)

Typical follow-up flow:

• download artifacts/scored_rows.jsonl.gz
• compute a new per-candidate reward from completion and any row context you need
• store the new score beside the existing meter/count scores
• recompute row-level summaries and difficulty labels from the rescored candidate set

This means generation is reusable: if you want to test a new reward later, you can usually start from the pushed completions rather than rerunning the full preprocess.

Local assembly summary

Machine-readable release summary:

{
  "assembled_at": "2026-04-11T07:22:16Z",
  "repo_id": "Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1",
  "run_dir": "/root/workspace/Shaer/grpo/outputs/preprocess_meter_count/full_train_k8_20260409_192820",
  "manifest_rows": 109070,
  "assembled_rows": 109070,
  "missing_scored_rows": 0,
  "train_jsonl": "/root/workspace/Shaer/grpo/outputs/preprocess_meter_count/full_train_k8_20260409_192820/assembled/train.jsonl",
  "difficulty_counts": {
    "easy": 33606,
    "hard": 29649,
    "medium": 45815
  },
  "base_meter_counts": {
    "البسيط": 16563,
    "الخفيف": 9263,
    "الرجز": 4311,
    "الرمل": 4316,
    "السريع": 6758,
    "الطويل": 26346,
    "الكامل": 20375,
    "المتقارب": 4877,
    "المجتث": 1938,
    "المديد": 504,
    "المنسرح": 2579,
    "الهزج": 573,
    "الوافر": 10667
  },
  "candidate_count_min": 8,
  "candidate_count_max": 8,
  "candidate_count_sum": 872560,
  "dry_run": false,
  "push": true,
  "include_completions": false,
  "scored_backup_requested": true,
  "dataset_rows": 109070,
  "uploaded_files": [
    "README.md",
    "artifacts/assembly_summary.json",
    "artifacts/run_config.json",
    "artifacts/model_meta.json",
    "artifacts/difficulty_rule.json",
    "artifacts/scored_rows.jsonl.gz"
  ],
  "scored_backup_bytes": 859750212
}