Read the dataset, checkpoint status, provenance and licensing notes before requesting access. This is an experimental research snapshot, not a production-qualified model.

This is Ampixa's 27 October 2026 research snapshot for a Nepali-English conversational receptionist based on Gemma 4 E4B. It contains fine-tuning data, the implementation history, evaluation summaries, preserved adapters/checkpoints, and reproduction records. Access is subject to manual approval.

This is a dataset and research repository, not a production-qualified model release. No checkpoint in this snapshot is claimed to pass all receptionist quality gates. Earlier pruning, drafting, action-head, serving and training experiments include failed candidates. Historical README entries and process logs describe their creation time, not a currently running job.

Dataset configuration	Train rows	Validation rows	Status
`kura_nepali_run3`	17,531	6,735	Exact published message/tool targets from the corpus used by completed Run 3; quality limitations remain
`balanced_transfer_seed`	112	None	24 semantic scenarios, four language views, 112 assistant targets; independently model-reviewed and tokenizer-checked, not native-human certified or trained

Historical curriculum versions are archived separately. They overlap heavily: do not concatenate them or sum their counts as independent examples. A dataset being included does not mean it was used in the final run or passed semantic/release gates. Read its original manifest.

The current full-vocabulary branch keeps 42 layers, hidden width 2560, FFN width 10240, vocabulary 262144 and the 4-bit affine/group-64 source weights. Training uses rank-8 Q/O LoRA on all 42 layers, with 3,325,952 trainable parameters. It makes no FFN, depth or vocabulary changes. Historical 32K-vocabulary branches and their adapters require their own matching compact base/tokenizer; do not attach them to the full-vocabulary base.

The original source is Anotos/Gemma-4-E4B-Heretic-MLX-4bit, revision 535cbb02a0a9360aaddb0a31b0353c195979d51e. Existing source behaviour is preserved; this project did not perform further refusal-behaviour editing.

Use the two dataset configurations with an approved Hugging Face token:

import json
from datasets import load_dataset

data = load_dataset("ampixa/e4b-finetune", "kura_nepali_run3", token=True)
row = data["train"][0]
messages = json.loads(row["messages_json"])
tools = json.loads(row["tools_json"])

The published table uses string columns for serialized messages and tools so differing tool argument schemas remain valid Arrow data. Decoding those columns recovers the original message and tool structures. Original complete rows, manifests and gold metadata remain in archives/training.tar.gz; message/tool content is not rewritten. Validation is development data, not an untouched test set. The balanced seed has no held-out split and must not be used to establish generalization.

Repository path	Contents
`data/`	Two directly loadable dataset configurations
`archives/project.tar.gz`	Source, tests, runbooks, experiment protocols and serving bridges
`archives/training.tar.gz`	Versioned authored curricula, admitted/candidate status and source manifests
`archives/fixtures.tar.gz`	Authored task fixtures; evaluation fixtures remain evaluation-only
`archives/reports.tar.gz`	Historical reports, aggregate metrics, reviews and test records selected without raw prompt/response journals
`archives/checkpoint_metadata.tar.gz`	Original checkpoint configuration, lineage and restoration metadata
`archives/metadata.tar.gz`	External-source/benchmark manifests whose raw data are not redistributed
`artifacts/`	Content-addressed preserved local weights; mappings in `metadata/checkpoint_blobs.json`
`checkpoints/`	Additional full-vocabulary Run 3 checkpoint copies from k2, when listed in the release manifest
`metadata/`	Per-file source hashes, inclusion/exclusion counts and published checksums

Restore the project snapshot into a new directory:

hf download ampixa/e4b-finetune --repo-type dataset --local-dir ./e4b-finetune
python ./e4b-finetune/restore_snapshot.py --snapshot ./e4b-finetune --output ./gemma4-ne-en-prune-restored
# Add --with-weights to restore the mapped local checkpoint blobs too.

This verifies archive and restored-file SHA-256 values. Runbooks preserve original host paths for provenance; pass your own explicit source-model/adapter/corpus paths when reproducing a run. Use the runtime versions in the corresponding manifest. The full original base is referenced by upstream revision, not duplicated; a historically pruned base is included only when listed in the checkpoint map. Additional checkpoints/ directories are already in loadable adapter layout and do not need archive extraction.

The primary workstreams represented are vocabulary/BPE and quantization-aware surgery; recovery curricula and retention failures; speculative/block drafting and action heads; MLX serving, cancellation and CUDA compatibility probes; multilingual tool/evidence evaluation; Nepali numeral, exact-entity and reading repair; and the new balanced-transfer experiment. No L40 capacity claim, 200-call production qualification, universal speedup, or successful sub-2GB model follows from this archive.

Recent controls completed 576 matched requests across original Heretic, Run 3 step 12000 and the final adapter. They reproduced two unusable malformed tool calls in the final adapter. Prompt-only changes did not reliably repair factual answers. The four balanced-transfer sampling schedules have been implemented, but their full training comparison has not started.

Raw customer calls, private review packets, credentials, source PDFs and unresolved mixed external QA payloads are excluded. The Orbit call-topic report is aggregate-only, uses overlapping pattern matches on 9,227 non-held-out calls, and is not certified primary-intent labeling. The archive includes outbound promotions as well as support calls. Exact exclusions and limitations are recorded rather than silently treating this as a byte-for-byte backup of every local cache.

See licensing and attribution, release inventory, source manifest, and published checksums. Original files remain on their source hosts. This published snapshot is on Hugging Face; older embedded provenance statements saying “NOT on HF” describe the earlier copy date and are superseded by this release's inventory.

Downloads last month: 24

Total file size:

3.24 GB