QwenGrad-DPO — M1 DPO calibration on selected M0-final-v2
Direct Preference Optimization applied to the healthy selected M0-final-v2 checkpoint of the OpenGrad tool-use study — not to the base model, and not to a historical DPO checkpoint.
Research artifact, not a production model.
This repository is published by Experimental Machines. It contains the promoted
dpo-checkpoint-30 at the repository root, in standard transformers layout.
| Original release | arrochi112/OpenGrad-Qwen3.5-2B-M1-DPO-CanonicalV2-Final-v2 at revision f33d20308982f37deb459076f489e794d5521ee3 (also holds checkpoints 60/90/120) |
| Base model | Qwen/Qwen3.5-2B at revision 15852e8c16360a2fea060d615a32b45270f8a8fc |
| Selection | checkpoint 30 of 120, frozen DEV balanced selector |
| Promotion | PROMOTED under the prospective tool_use_promotion_v4 policy |
Results (pre-registered internal confirmatory partition, 1,277 examples)
The chart reports the exact values for call_f1, precision, recall, over_call, clarification,
and unsupported. Higher is better for every metric except over_call.
M1 preserves the M0 calibrated frontier and makes a small improvement in call F1 and recall. It does not materially reduce over-calling; this is calibration retention and slight improvement, not a large frontier movement.
Evaluation and promotion
Checkpoint selection used the frozen DEV partition (2,373 examples, fingerprint 88a56821…).
Checkpoints 30/60 were within the pre-registered 0.01 macro tolerance, so the earlier checkpoint 30
was selected. The confirmatory partition (fingerprint d6d1e394…) was then scored exactly once on
checkpoint 30.
The prospective tool_use_promotion_v4 policy passed: precision, recall, and macro floors; over-call
ceiling; clarification and unsupported floors; parse validity; and M0-relative regression checks.
Checkpoint 30 is PROMOTED. This policy does not compare recall to B0's degenerate always-call
recall.
The confirmatory partition is pre-registered internal evidence, not an untouched external test.
The evaluation population has no ANSWER examples, so no_call_accuracy is NA. Tool-selection
accuracy, argument validity, and schema validity are not computed by the current evaluator and are
not treated as satisfied.
Frozen lineage
- Parent experiment:
m0_sft_canonical_v2_final - Parent checkpoint:
checkpoint-1800 - Parent model hash:
7144579aeecec8b4de25f193ab63085efdf8d9d76b85ed915352291b0152277a - Preference dataset: 481 local calibration pairs, hash
d39168948d09fc3c355cd83f9f0857f310086322b0968fd2e7d78125150faef4(m1_calibration_preference_pairs_v1) - DPO: beta 0.05, learning rate 5e-7, cosine schedule, 120 steps, seed 42, bfloat16
- Training commit:
bb15c7e0181c7f526875cd1426672431ae14bd83
The preference set combines deterministic base/M0 disagreements on Canonical-v2 training prompts with a bounded curated When2Call training slice. Frozen behavioral evaluation IDs were excluded; no paid external API was used. All pair origins and input hashes are recorded in the OpenGrad repository.
Intended use
Research artifact. Not safety-tuned, not aligned, and not intended for autonomous tool use. It inherits the limitations of its public training sources and of its evaluator's unmeasured dimensions. The base model's own license and terms continue to apply to these weights.
Provenance
The model card text above mirrors the original OpenGrad release card for this checkpoint; only the repository-level framing (mirror, Experimental Machines) has been added. The evidence behind every number is in the OpenGrad repository:
- Execution:
reports/M1_DPO_EXECUTION_REPORT.md - Evaluation:
reports/M1_DPO_EVALUATION.md - Machine-readable status:
docs/EXPERIMENT_STATUS.md - Next-stage decision:
reports/M2_DECISION.md
Published by Experimental Machines.