Document training status and release requirements

README.md

@@ -1,3 +1,95 @@
 ---
-license: apache-2.0
+license: "apache-2.0"
+language:
+- en
+- zh
+tags:
+- physics
+- research
+- experimental
+- training-in-progress
 ---
+
+# HyperSol/ResearchPhysics-27B
+
+**训练中 / Training in progress — experimental, unreleased.**
+
+Status updated: 2026-10-29. This is the maintainer-declared project stage. It does
+not attest that a GPU job is currently running, that training has completed, or
+that a checkpoint has passed evaluation. This repository is not a validated
+model release. Existing files, if any, have not been certified by this card.
+
+## Model description and scope
+
+ResearchPhysics studies whether research-paper-derived problems involving
+interacting physical mechanisms and constraints can improve multi-step reasoning
+on unseen papers and tasks. This repository is reserved for the 27B research track. The main planned comparison uses matched Qwen3.5-27B Base/SFT with thinking enabled; this plan does not identify the base of an uploaded checkpoint. The repository name identifies the intended
+experiment family; it does not establish the architecture or parameter count of
+any files present here. No capability or improvement claim is made at this stage.
+
+Intended use: controlled research after a specific checkpoint and its provenance
+have been verified. Outputs can contain incorrect equations, assumptions,
+citations, or conclusions and require independent scientific checking.
+
+## Provenance required for a release
+
+| Field | Current disclosure |
+| --- | --- |
+| Exact base model and immutable revision | Pending checkpoint-specific verification |
+| Architecture, parameter count and tokenizer | Pending checkpoint/config verification |
+| Training method and run ID | Supervised fine-tuning (SFT), intended main comparison; exact run binding pending |
+| Training code commit and dependency versions | Pending immutable release manifest |
+| Training data snapshot, revision and SHA256 | Private; exact checkpoint-to-snapshot binding pending |
+| Hyperparameters, thinking/template settings, masking and seeds | Pending actual executed configuration |
+| Hardware, duration and training completion | Not certified by this repository card |
+| Checkpoint step/epoch and weight hashes | Pending verified export |
+| Base-model and training-data usage terms | Pending checkpoint-specific license review |
+
+The private data repository for this track is [HyperSol/ResearchPhysics-SFT](https://huggingface.co/datasets/HyperSol/ResearchPhysics-SFT).
+It is a separate access-controlled repository. A link does not imply that a
+snapshot has been uploaded, that access is granted, or that the data may be
+redistributed. Supervision sources and formatting variables must be disclosed
+for each released checkpoint.
+
+## Evaluation and evidence
+
+No independently verified results are published in this card. A future result
+must identify the exact Base/SFT/RL checkpoint, held-out split and contamination
+controls, evaluation version, thinking/token budget, generation settings, scoring
+protocol, sample count, uncertainty and failure cases. Comparisons must use
+matched settings and distinguish training-set observations from held-out
+benchmarks. Missing scores are unscored, not zero.
+
+Training loss, successful export, or readable model files do not establish
+scientific validity, generalization, or a causal training benefit. Historical
+9B explorations must not be pooled with the matched 27B Base/SFT main comparison.
+
+## Limitations, risks and permitted interpretation
+
+Physics reasoning can fail through invalid approximations, omitted boundary
+conditions, dimensional errors or incomplete derivations. Answer fluency does
+not establish correctness. Robustness, memorization, benchmark contamination,
+out-of-domain behavior and safety have not been established for a release here.
+The model is not validated for consequential engineering or scientific decisions.
+
+## Access, licensing and availability
+
+The model repository is public; training datasets remain private. Public
+visibility does not grant access to private data or replace applicable upstream
+licenses. Existing license metadata has been retained; its applicability to a future checkpoint still requires verification.
+
+No supported inference command is provided until tokenizer/template compatibility
+and a specific exported checkpoint have been verified. Do not infer download
+readiness from the repository name or this status card.
+
+## Release checklist
+
+- [ ] Pin the base model, code, executed configuration and data snapshot.
+- [ ] Verify exported weights, tokenizer/template and checkpoint hashes.
+- [ ] Complete checkpoint-specific license and data-use review.
+- [ ] Publish a matched independent evaluation with failure analysis.
+- [ ] Replace this training-status card with a versioned release card.
+
+This card does not alter existing weights. Earlier cards remain in the
+repository commit history. Project stage and release claims must be updated
+from evidence rather than inferred from an upload.