Allanatrix/nexa-llama3-8b-science-multitask-merged

Nexa Llama-3 8B Science Multitask (Merged)

Merged full model produced by fusing LoRA adapters trained for scientific multitask instruction tuning.

Model Details

• Base model: meta-llama/Meta-Llama-3-8B
• Method: QLoRA/LoRA adapter training, then merged (merge_and_unload) into full weights
• Timestamp (UTC): 2026-02-24T03:56:05+00:00

Tasks

• <TASK:VERIFY>: SUPPORTS/REFUTES/NEI claim verification
• <TASK:QA>: yes/no/maybe abstract-grounded QA
• <TASK:RERANK>: 0-3 relevance scoring used for ranking

Training Data

• Dataset: Nexa science multitask mixture (balanced short rerun release)
• Format: text-to-text with explicit task tokens and JSON outputs

Evaluation Snapshot

Balanced split (trusted)

Metric	Baseline (pre-rerun)	Post-train
Verify Accuracy	0.5333	0.6667
Verify Macro-F1	0.5385	0.6592
QA Accuracy	0.4000	0.5333
QA Majority Baseline	0.4000	0.4000
Rerank Pair Accuracy	0.3500	0.4667
Rerank MRR@10	0.2667	0.5708
Rerank Recall@1	0.0000	0.5000
Rerank Recall@3	0.3333	0.5000
Rerank Recall@5	0.5000	0.6667

Mixed split (diagnostic only)

• Verify Accuracy: 0.5833
• Verify Macro-F1: 0.6667
• QA Accuracy: 0.6667 (mixed split is label-skewed)
• Rerank MRR@10: 0.4352

Intended Use

Research and prototyping for scientific assistant workflows that mix verification, QA, and reranking.

Limitations

• Biomedical/scientific outputs can still hallucinate or overstate confidence.
• Not validated for clinical, legal, or high-stakes decision making.
• Mixed validation split has known QA label imbalance and should not be used as sole quality signal.

Artifacts in This Repo

• Merged model weights and tokenizer
• eval/ metrics JSON files
• code/ dataset/training/eval scripts used in this release

Notes

Merged from Nexa_Tune_Balanced_Rerun adapter after balanced short rerun.

HF repo: https://huggingface.co/Allanatrix/nexa-llama3-8b-science-multitask-merged