roblee100/qwen2.5-32b-haem-myeloid-parser-v3-atomic

Qwen2.5 32B Haem Myeloid Parser v3 Atomic Adapter

This is a PEFT/LoRA adapter trained for haem.io atomic myeloid genetics and cytogenetics parser tasks. It is intended to be loaded with the base model Qwen/Qwen2.5-32B-Instruct or the locally mirrored equivalent used during testing (/root/hf-models/Qwen2.5-32B-Instruct).

Intended Use

The adapter extracts atomic structured evidence for the haem.io parser, especially:

• AML recurrent genetic abnormality flags
• MDS mutation flags
• MDS cytogenetic flags
• TP53 evidence decomposition
• CEBPA evidence decomposition
• raw karyotype extraction

Morphology and blast percentages are intentionally out of scope; those are expected to be entered manually in the GUI.

This adapter should be used with the deterministic haem.io parser guardrails. It should not be treated as an unguarded raw drop-in replacement.

Training Run

• Base model: Qwen2.5-32B-Instruct
• Method: QLoRA / PEFT adapter
• Adapter output: adapter-final
• Training output dir on RunPod: /root/qwen32b-parser-v3-atomic
• Final eval loss from training: see training/eval_metrics.json

Eval Summary

Parser Adapter Eval Results

Eval	Total	Valid JSON	Schema Valid	Exact	Max Tokens
qwen32b-parser-v3-atomic/eval-heldout	72	72/72 (100.0%)	72/72 (100.0%)	61/72 (84.7%)	1024
qwen32b-parser-v3-atomic/eval-report-1024	42	41/42 (97.6%)	41/42 (97.6%)	30/42 (71.4%)	1024
qwen32b-parser-v3-atomic/eval-report-row33-2048	1	1/1 (100.0%)	1/1 (100.0%)	0/1 (0.0%)	2048
qwen32b-parser-v3-atomic/eval-adversarial-report-2048	72	72/72 (100.0%)	72/72 (100.0%)	62/72 (86.1%)	2048
qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048	10	10/10 (100.0%)	10/10 (100.0%)	0/10 (0.0%)	2048
qwen32b-parser-v3-atomic/eval-runx1-fusion-prompt-v2-2048	1	1/1 (100.0%)	1/1 (100.0%)	1/1 (100.0%)	2048

By Task

qwen32b-parser-v3-atomic/eval-heldout

Task	Total	Valid JSON	Schema Valid	Exact
aml_recurrent_flags	14	14	14	14
cebpa_evidence	7	7	7	3
karyotype_extraction	11	11	11	11
mds_cytogenetics_flags	12	12	12	7
mds_mutation_flags	16	16	16	16
tp53_evidence	12	12	12	10

qwen32b-parser-v3-atomic/eval-report-1024

Task	Total	Valid JSON	Schema Valid	Exact
aml_recurrent_flags	7	7	7	3
cebpa_evidence	7	7	7	6
karyotype_extraction	7	7	7	7
mds_cytogenetics_flags	7	6	6	5
mds_mutation_flags	7	7	7	3
tp53_evidence	7	7	7	6

qwen32b-parser-v3-atomic/eval-report-row33-2048

Task	Total	Valid JSON	Schema Valid	Exact
mds_cytogenetics_flags	1	1	1	0

qwen32b-parser-v3-atomic/eval-adversarial-report-2048

Task	Total	Valid JSON	Schema Valid	Exact
aml_recurrent_flags	12	12	12	11
cebpa_evidence	12	12	12	11
karyotype_extraction	12	12	12	12
mds_cytogenetics_flags	12	12	12	7
mds_mutation_flags	12	12	12	11
tp53_evidence	12	12	12	10

qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048

Task	Total	Valid JSON	Schema Valid	Exact
aml_recurrent_flags	1	1	1	0
cebpa_evidence	1	1	1	0
mds_cytogenetics_flags	5	5	5	0
mds_mutation_flags	1	1	1	0
tp53_evidence	2	2	2	0

qwen32b-parser-v3-atomic/eval-runx1-fusion-prompt-v2-2048

Task	Total	Valid JSON	Schema Valid	Exact
mds_mutation_flags	1	1	1	1

Failure Digest

• qwen32b-parser-v3-atomic/eval-heldout row 7 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 18 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 21 cebpa_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 30 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 33 cebpa_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 41 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 44 cebpa_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 52 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 55 cebpa_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 64 tp53_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-heldout row 70 tp53_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 1 aml_recurrent_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 5 cebpa_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 7 aml_recurrent_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 13 aml_recurrent_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 19 aml_recurrent_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 20 mds_mutation_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 26 mds_mutation_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 27 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 32 mds_mutation_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 33 mds_cytogenetics_flags: valid_json=False schema_valid=False error=JSONDecodeError: Unterminated string starting at: line 1 column 1875 (char 1874)
• qwen32b-parser-v3-atomic/eval-report-1024 row 34 tp53_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-1024 row 38 mds_mutation_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-report-row33-2048 row 1 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 3 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 9 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 28 tp53_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 40 tp53_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 47 cebpa_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 49 aml_recurrent_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 50 mds_mutation_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 51 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 57 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-report-2048 row 63 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 1 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 2 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 3 tp53_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 4 tp53_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 5 cebpa_evidence: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 6 aml_recurrent_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 7 mds_mutation_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 8 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 9 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None
• qwen32b-parser-v3-atomic/eval-adversarial-focused-prompt-examples-2048 row 10 mds_cytogenetics_flags: valid_json=True schema_valid=True error=None

Notes

See SEMANTIC_EVAL_NOTES.md for interpretation of exact-match failures versus clinically meaningful field behavior. Raw generations and summaries are stored under evals/.