| """Publish the measured v2 model page after the merged checkpoint passes.""" |
| import json,hashlib |
| from huggingface_hub import HfApi,hf_hub_download,CommitOperationAdd |
| from pathlib import Path |
| api=HfApi();repo="SEE7/Astheral" |
| def read(name):return Path(hf_hub_download(repo,name)).read_text() |
| m=json.loads(read("evaluation.json")) |
| assert m["version"]=="2" and m["quality_gate_passed"] |
| base=m["base_scores"];new=m["merged_scores"] |
| assert new["regression"]["passed"]==2 |
| assert new["number"]["passed"]>=base["number"]["passed"] |
| assert new["fresh"]["passed"]==new["fresh"]["total"] |
| assert m["mode_adherence"]["passed"]==m["mode_adherence"]["total"] |
| def fraction(s,key):return str(s[key]["passed"])+"/"+str(s[key]["total"]) |
| def formats(s):return str(s["exact"]["passed"]+s["json"]["passed"])+"/"+str(s["exact"]["total"]+s["json"]["total"]) |
| page="""--- |
| language: |
| - en |
| license: llama3.1 |
| library_name: transformers |
| pipeline_tag: text-generation |
| base_model: NousResearch/DeepHermes-3-Llama-3-8B-Preview |
| base_model_relation: finetune |
| model_name: Llama-3.1-Astheral |
| tags: |
| - astheral |
| - llama |
| - instruct |
| - reasoning |
| - qlora |
| --- |
| |
| # Astheral |
| |
| **Built with Llama.** Astheral is SEE7's 8.03B assistant for instruction following, coding, concise explanations and fictional roleplay. |
| |
| ## Instruct and thinking |
| |
| Choose **instruct** for direct answers or **thinking** for a worked solution before the final answer. Both modes use the same trained weights. The included inference script selects the mode and accepts your own system instructions with --system. |
| |
| ## What's changed in v2 |
| |
| - Balanced training across general instruction tasks, formatting, text editing, Python, and multi-step mathematics. |
| - Rank-32 adapters across attention and feed-forward projections. |
| - Explicit training to account for all groups in counting problems and preserve requested output structure. |
| - Short, complete worked solutions in thinking mode. |
| - A saved, merged BF16 checkpoint plus the separately downloadable adapter. |
| |
| """ |
| page+=f"Training used **{m['training_examples']:,} examples**, **{m['optimizer_steps']} optimizer steps**, and **{m['trainable_parameters']:,} trainable adapter parameters** on a remote **{m['hardware']}**. The original DeepHermes weights provide the starting model.\n\n" |
| page+="## Measured results\n\nBoth models used identical prompts and deterministic generation settings: up to 384 new tokens for instruct and 1,024 for thinking. The core suite contains 14 formatting tasks, 16 numerical reasoning tasks and six executable Python tasks.\n\n" |
| page+="| Check | DeepHermes base | Astheral v2 |\n|---|---:|---:|\n" |
| page+=f"| Formatting | {formats(base)} | {formats(new)} |\n| Numerical reasoning | {fraction(base,'number')} | {fraction(new,'number')} |\n| Executable Python | {fraction(base,'code')} | {fraction(new,'code')} |\n| Core checks total | {fraction(base,'fresh')} | {fraction(new,'fresh')} |\n| Known regression cases | {fraction(base,'regression')} | {fraction(new,'regression')} |\n\n" |
| page+="The two regression cases cover the alpha/omega line format and counting the remaining red and blue balls together. Full prompts, answers, checker code and run metrics are available in evaluate.py, base_outputs.json, astheral_merged_outputs.json and evaluation.json. Six additional samples cover roleplay, prompt design, authorized networking explanations, uncertainty, debugging and summarization.\n\n" |
| page+=f"Default mode formatting was verified in **{m['mode_adherence']['passed']}/{m['mode_adherence']['total']} responses**: instruct responses omit thinking tags; thinking responses contain a completed thinking block.\n\n" |
| if "confirmation" in new: |
| assert new["confirmation"]["passed"]==new["confirmation"]["total"] |
| page+=f"The saved checkpoint also passed **{new['confirmation']['passed']}/{new['confirmation']['total']} additional confirmation cases** covering new CSV records, literal copying, fraction-based balances, time conversion and complete Python functions. Their prompts and answers are in confirmation_outputs.json.\n\n" |
| if any(item.rfilename=="holdout_evaluation.json" for item in api.model_info(repo).siblings): |
| h=json.loads(read("holdout_evaluation.json")) |
| assert h["no_material_regression_gate"] |
| page+="## Independent test subsets\n\n| Test subset | DeepHermes base | Astheral v2 |\n|---|---:|---:|\n" |
| for key,label in [("number","GSM8K test problems"),("code","HumanEval functions")]: |
| old=h["scores"]["base"][key];new_h=h["scores"]["astheral"][key] |
| page+=f"| {label} | {old['passed']}/{old['total']} | {new_h['passed']}/{new_h['total']} |\n" |
| page+="\nExact task IDs, generation limits, source revisions, outputs and executable tests are in holdout_evaluation.json, holdout_outputs.json and holdout_eval.py.\n\n" |
| page+="## Run on a remote GPU\n\nInstall a CUDA-enabled PyTorch build and requirements.txt, then download inference.py from this repository.\n\n" |
| page+=" python inference.py \"Explain Python generators.\" --mode instruct\n python inference.py \"A box has 23 red and 17 blue beads; 6 red are removed. How many remain?\" --mode thinking\n python inference.py \"Write a fictional starship announcement.\" --mode instruct --load-in-4bit\n\n" |
| page+="Use --system \"Your instructions\" to replace the default assistant instructions. Decoding is deterministic by default; add --sample for creative variation. BF16 weights occupy about 16 GB; --load-in-4bit loads the weights in NF4 for lower GPU memory use. The script loads SEE7/Astheral directly. Dockerfile includes the CUDA runtime and compiler dependencies for the 4-bit backend.\n\n" |
| if any(item.rfilename=="cli_smoke_test.json" for item in api.model_info(repo).siblings): |
| cli=json.loads(read("cli_smoke_test.json"));assert cli["all_passed"] |
| page+="The published inference script was executed on a remote GPU and verified in instruct mode, thinking mode and with a replacement system instruction. The actual arguments and generated answers are recorded in cli_smoke_test.json.\n\n" |
| if api.repo_exists("SEE7/Astheral-NF4"): |
| compact=json.loads(Path(hf_hub_download("SEE7/Astheral-NF4","evaluation.json")).read_text()) |
| assert compact["quality_gate_passed"] and compact["saved_checkpoint_cli_verified"] |
| page+=f"## Compact download\n\n[Astheral NF4](https://huggingface.co/SEE7/Astheral-NF4) provides a prequantized **{compact['weight_bytes']/1e9:.2f} GB** checkpoint. It was evaluated separately and reloaded through the inference script in both modes. Its complete results are included in that repository.\n\n" |
| page+="## Files and reproducibility\n\nThe repository includes merged Safetensors weights, the adapter folder, train.py, prepare_data.py, the exact training and validation examples, trainer logs, data source revisions, output comparisons and artifact hashes. The merged checkpoint was saved, reloaded, and evaluated before release.\n\n" |
| page+="## Training data and attribution\n\nTraining combines original Astheral examples with selected components of HuggingFaceTB/smoltalk (Apache 2.0), the openai/gsm8k training split (MIT), and bigcode/self-oss-instruct-sc2-exec-filter-50k (ODC Attribution). See data_sources.json, DATA.md, DATA_LICENSE files and NOTICE.\n\n" |
| page+="Astheral is fine-tuned from NousResearch/DeepHermes-3-Llama-3-8B-Preview, revision 53d902c7dcfa4c749a3b455149064f82af2e2549, built on Meta Llama 3.1 8B. The project is Astheral; the distributed model name is Llama-3.1-Astheral. See LICENSE and UPSTREAM_LLAMA3_LICENSE.\n" |
| manifest=json.loads(read("artifact_manifest.json")) |
| |
| generation=json.loads(read("generation_config.json")) |
| generation.update(do_sample=False,temperature=1.0,top_p=1.0,top_k=50,max_new_tokens=1024) |
| inference=Path(__file__).with_name("inference.py").read_text() |
| compile(inference,"inference.py","exec") |
| docker=Path(__file__).with_name("Dockerfile").read_text() |
| changed={"README.md":page,"inference.py":inference,"Dockerfile":docker,"generation_config.json":json.dumps(generation,indent=2)} |
| for name,body in changed.items():manifest[name]={"bytes":len(body.encode()),"sha256":hashlib.sha256(body.encode()).hexdigest()} |
| changed["artifact_manifest.json"]=json.dumps(manifest,indent=2) |
| api.create_commit(repo_id=repo,operations=[CommitOperationAdd(path_in_repo=name,path_or_fileobj=body.encode()) for name,body in changed.items()],commit_message="Present Astheral v2 capabilities, measured comparisons and deterministic inference defaults") |
| info=api.model_info(repo,files_metadata=True) |
| assert not info.private |
| if any(tag.name=="v2" for tag in api.list_repo_refs(repo).tags):api.delete_tag(repo_id=repo,tag="v2") |
| api.create_tag(repo_id=repo,tag="v2",revision=info.sha,exist_ok=True) |
| assert next(tag for tag in api.list_repo_refs(repo).tags if tag.name=="v2").target_commit==info.sha |
| print("RELEASE_VERIFIED",json.dumps({"revision":info.sha,"base":base,"astheral":new,"training_examples":m["training_examples"],"trainable_parameters":m["trainable_parameters"],"hardware":m["hardware"]}),flush=True) |
|
|
