efe-T/gemma4-e4b-efe-t-checkpoints

Gemma 4 E4B IT — Efe-T adapter and checkpoints

Experimental language LoRA adapter trained for one epoch on the deduplicated public efe-T dataset collection. This repository contains adapters and resumable checkpoints; it is not a standalone full Transformers model at the repository root.

Saved training result

Final adapter: checkpoints/checkpoint-118.

The final checkpoint includes adapter weights/config, tokenizer/chat template, optimizer, scheduler, RNG, gradient scaler, and Trainer state. Training completed 118 optimizer steps, one epoch. A local backup of all 13 checkpoint files was checked against remote file sizes; all 7 LFS file SHA-256 hashes matched. Adapter payload length and PyTorch state ZIP integrity checks passed.

Base model revision: 6f10b2aa522ba471321d8246fd71a3dfa8c5b9a2. Official Google chat-template revision: ee0ef6023621cff504d758262d4e04895a5af4a2. Benchmark adapter snapshot: repository revision 2a20e535a23afe8103d18c93fe7cb6ec21ef22c6, subfolder checkpoints/checkpoint-118.

Quick benchmark results — 18 October 2026

32 fixed, seeded examples per benchmark. These are sampled checks, not full benchmark or leaderboard results. Original quantized base and final adapter used the same examples, tokenizer, prompts, and scoring. Seed 3407; zero-shot raw continuation likelihood without a chat wrapper or generated reasoning. Task formatting and character-length-normalized scoring follow the EleutherAI lm-evaluation-harness ARC and HellaSwag task definitions.

Benchmark	N	Original accuracy	Adapter accuracy	Original acc_norm	Adapter acc_norm
ARC-Challenge, test	32	15/32 (46.875%)	16/32 (50.000%)	17/32 (53.125%)	17/32 (53.125%)
HellaSwag, validation	32	18/32 (56.250%)	18/32 (56.250%)	22/32 (68.750%)	22/32 (68.750%)

acc_norm chooses the answer with the largest continuation log likelihood divided by answer character length. Aggregate normalized scores were unchanged; raw ARC accuracy improved by one question. There is substantial sampling uncertainty. No benchmark contamination audit was performed. These results do not establish broad improvement, absence of forgetting, or thinking quality. They evaluate the pre-export quantized Transformers base with/without the adapter; they are not measurements of any later GGUF export.

Reproducible artifacts:

• Report
• Summary and pinned dataset revisions/sample indices
• Per-example predictions and likelihoods
• Configuration

Task definitions: ARC, HellaSwag.

Training and internal validation

4-bit QLoRA; frozen original base parameters; rank 8, alpha 8, dropout 0; 18,350,080 trainable language attention/MLP adapter parameters. AdamW 8-bit, peak LR 1e-5, cosine decay, warmup ratio 0.1, weight decay 0.01, clipping 0.3. Microbatch 1, accumulation 16, context limit 2048, seed 3407. Single Tesla T4; Unsloth 2026.10.6, Transformers 5.5.0, Torch 2.11.0+cu128.

Internal evaluation	Before training	Final
Validation loss (first 64 held-out rows)	3.757906	3.057740
Retention reference-answer loss (16 basic probes)	3.861217	2.925237

Loss reductions are not generation accuracy. The retention probes are a small heuristic set, not a comprehensive capability benchmark. No validation-based early stopping was used; nonfinite metrics or sufficiently worsened retention loss stop the run. No automatic deployment occurred.

Data and limitations

All 12 public efe-T datasets represented. Raw 2,566 conversations became 1,886 train / 300 validation / 300 test after removing 80 duplicates. All distinct training rows were used; test rows remained unused. Only the final assistant response was supervised, with earlier assistant turns and padding masked. Existing separate reasoning annotations were preserved (116 training rows).

About 63% of training rows come from a template-heavy starter dataset. Numeric variants of two task patterns occur across splits. No exact cross-split conversation duplicates were found, but there was no complete semantic near-duplicate audit. Prose and translations were not all independently verified. All 2,486 conversations passed template/prefix checks; maximum length 1,353 tokens; 2,599 selected deterministic assertions passed.

Suitable as an experimental adaptation candidate. Review outputs on your intended tasks before use.