Hon-Wong's picture
Release 795,558 execution-verified L3 examples
8b2b426 verified
|
Raw History Blame Contribute Delete
5.48 kB
---
pretty_name: LANA Code L3 Verified
license: apache-2.0
language:
- en
task_categories:
- text-generation
tags:
- code
- synthetic
- reasoning
- execution-verified
- ultradata-code
- sft
size_categories:
- 100K<n<1M
configs:
- config_name: all
default: true
data_files:
- split: train
path: data/*/train-*.parquet
- config_name: main_32k
data_files:
- split: train
path: data/main_32k/train-*.parquet
- config_name: recovery_64k
data_files:
- split: train
path: data/recovery_64k/train-*.parquet
extra_gated_prompt: Access requests are reviewed manually by LANA-research. Please describe your intended use.
extra_gated_fields:
Affiliation: text
Intended use: text
extra_gated_button_content: Request access
---
# LANA Code L3 Verified
A frozen release of **795,558** Python/C++ training examples derived from [OpenBMB UltraData-Code L3](https://huggingface.co/datasets/openbmb/UltraData-Code), regenerated with **Qwen3.8-Flash-Next** and execution-filtered using independently repaired tests. Snapshot prepared on 2026-10-23 (Europe/Rome).
The repository is **public and gated with manual approval**. This is a verified admitted subset, **not the entirety of L3**. Unresolved execution failures, benchmark-overlap holds, incomplete generations and length-rejected records are not in this release. The existing production training manifest was not modified by publishing it.
## Contents
| Config | Rows | Description |
|---|---:|---|
| `all` | 795,558 | Union of the two non-overlapping cohorts below |
| `main_32k` | 789,311 | Previously admitted records fitting the recorded 32K native packing criterion |
| `recovery_64k` | 6,247 | Newly recovered records; includes long examples requiring up to 64K |
Languages: `{"cpp": 34768, "py": 760790}`. Native serialized tokens: **2,291,403,507** under our recorded SLF/Nemotron template/tokenizer protocol, not the Qwen teacher tokenizer. Context eligibility follows `prefix_tokens + 2 * target_tokens + 1024 <= context_length`; do not infer eligibility from plain text length. The main cohort can contain fewer examples eligible for 16K; consult each row's `native.fullfit`. No answers are truncated. The `all` config already includes `main_32k` and `recovery_64k`; do not concatenate all three configs.
## Schema
- `uuid`: unique generated record identifier.
- `lang`: `py` or `cpp`.
- `messages`: the complete user question and assistant answer, preserving recorded reasoning and response text.
- `record_json`: a JSON string containing problem, executable solution, repaired test, original test (historical input, which may have been incorrect), native token lengths/context eligibility, source identifiers, hashes and execution evidence. **Use `repaired_test` for the recorded verification, not `original_test`.** Internal cluster paths are reduced to artifact basenames and internal service endpoints are omitted; content fields remain unchanged.
## Construction and verification
1. Independent task-only answer generation sends the question without a system prompt or reference answer. Original source solutions are not substituted as assistant training targets.
2. Test repair sees the task and original tests, never the candidate solution. Original tests can contain errors.
3. Candidate solutions execute against repaired tests in bounded environments. Admission requires recorded witness evidence of multiple distinct candidate-reaching assertion sites, not merely process exit success. Original-solution checks and prior tier metadata are preserved where available; missing original-solution evidence is not fabricated.
4. Question, original source-body and tested solution keys are checked for duplicates; a repeated key excludes a record. A further cross-cohort exact-hash check was performed for this release. These are exact-key checks, not a claim of complete semantic deduplication.
5. The inherited pipeline applies benchmark-overlap checks; this is not a guarantee that every semantic or translated benchmark overlap was detected.
**Passing generated/repaired tests is evidence of correctness on those tests, not a formal proof of general correctness.** Teacher mistakes, underspecified problems, limited tests and parser mistakes can remain. `formal_correctness_proven` is false. Main and recovery cohorts preserve their per-example quality tiers and evidence.
## Use
```python
from datasets import load_dataset
import json
ds = load_dataset("LANA-research/LANA-Code-L3-Verified", "all", split="train", token=True)
row = ds[0]
record = json.loads(row["record_json"])
messages = row["messages"]
```
An individually approved Hugging Face account is required to download the data. This is a training-only release; it has no newly created validation/test split. Do not use it as an unbiased evaluation benchmark.
## Provenance and license
Derived from OpenBMB UltraData-Code L3, released under Apache License 2.0. The inherited attribution and license are retained in `NOTICE` and `LICENSE`. LANA-research contributes regenerated answers, repaired-test processing, execution-filtered selection and packaging. Qwen3.8-Flash-Next is the generation teacher; its model licensing is distinct from the upstream dataset license. No endorsement by OpenBMB or Qwen is implied.
`release_manifest.json` records retained counts, native token totals, file sizes and SHA-256 digests. Counts describe this frozen release; ongoing production may subsequently yield more records.